Merge branch 'add_versatile_diffusers' of https://github.com/huggingface/diffusers into add_versatile_diffusers

docs/source/_toctree.yml

@@ -110,6 +110,8 @@
       title: "Stochastic Karras VE"
     - local: api/pipelines/dance_diffusion
       title: "Dance Diffusion"
+    - local: api/pipelines/versatile_diffusion
+      title: "Versatile Diffusion"
     - local: api/pipelines/vq_diffusion
       title: "VQ Diffusion"
     - local: api/pipelines/repaint

docs/source/api/pipelines/versatile_diffusion.mdx

@@ -0,0 +1,82 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# VersatileDiffusion
+
+VersatileDiffusion was proposed in [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) by Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, Humphrey Shi .
+
+The abstract of the paper is the following:
+
+*The recent advances in diffusion models have set an impressive milestone in many generation tasks. Trending works such as DALL-E2, Imagen, and Stable Diffusion have attracted great interest in academia and industry. Despite the rapid landscape changes, recent new approaches focus on extensions and performance rather than capacity, thus requiring separate models for separate tasks. In this work, we expand the existing single-flow diffusion pipeline into a multi-flow network, dubbed Versatile Diffusion (VD), that handles text-to-image, image-to-text, image-variation, and text-variation in one unified model. Moreover, we generalize VD to a unified multi-flow multimodal diffusion framework with grouped layers, swappable streams, and other propositions that can process modalities beyond images and text. Through our experiments, we demonstrate that VD and its underlying framework have the following merits: a) VD handles all subtasks with competitive quality; b) VD initiates novel extensions and applications such as disentanglement of style and semantic, image-text dual-guided generation, etc.; c) Through these experiments and applications, VD provides more semantic insights of the generated outputs.*
+
+*Overview*:
+
+| Pipeline | Tasks | Colab | Demo
+|---|---|:---:|:---:|
+| [pipeline_alt_diffusion.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py) | *Text-to-Image Generation* | - | -
+| [pipeline_alt_diffusion_img2img.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py) | *Image-to-Image Text-Guided Generation* | - |-
+
+## Tips
+
+- VersatileDiffusion is conceptually very similar as [Stable Diffusion](./api/pipelines/stable_diffusion),  but instead of providing just a image data stream conditioned on text, VersatileDiffusion provides both a image and text data stream and can be conditioned on both text and image.
+
+- *Run VersatileDiffusion*
+
+All task VersatileDiffusion can be tested very easily with the [`VersatileDiffusionPipeline`], [`VersatileDiffusionImg2ImgPipeline`] and the `"BAAI/VersatileDiffusion-m9"` checkpoint exactly in the same way it is shown in the [Conditional Image Generation Guide](./using-diffusers/conditional_image_generation) and the [Image-to-Image Generation Guide](./using-diffusers/img2img).
+
+- *How to load and use different schedulers.*
+
+The alt diffusion pipeline uses [`DDIMScheduler`] scheduler by default. But `diffusers` provides many other schedulers that can be used with the alt diffusion pipeline such as [`PNDMScheduler`], [`LMSDiscreteScheduler`], [`EulerDiscreteScheduler`], [`EulerAncestralDiscreteScheduler`] etc.
+To use a different scheduler, you can either change it via the [`ConfigMixin.from_config`] method or pass the `scheduler` argument to the `from_pretrained` method of the pipeline. For example, to use the [`EulerDiscreteScheduler`], you can do the following:
+
+```python
+>>> from diffusers import VersatileDiffusionPipeline, EulerDiscreteScheduler
+
+>>> pipeline = VersatileDiffusionPipeline.from_pretrained("BAAI/VersatileDiffusion-m9")
+>>> pipeline.scheduler = EulerDiscreteScheduler.from_config(pipeline.scheduler.config)
+
+>>> # or
+>>> euler_scheduler = EulerDiscreteScheduler.from_pretrained("BAAI/VersatileDiffusion-m9", subfolder="scheduler")
+>>> pipeline = VersatileDiffusionPipeline.from_pretrained("BAAI/VersatileDiffusion-m9", scheduler=euler_scheduler)
+```
+
+
+- *How to conver all use cases with multiple or single pipeline*
+
+If you want to use all possible use cases in a single `DiffusionPipeline` we recommend using the `components` functionality to instantiate all components in the most memory-efficient way:
+
+```python
+>>> from diffusers import (
+...     VersatileDiffusionPipeline,
+...     VersatileDiffusionImg2ImgPipeline,
+... )
+
+>>> text2img = VersatileDiffusionPipeline.from_pretrained("BAAI/VersatileDiffusion-m9")
+>>> img2img = VersatileDiffusionImg2ImgPipeline(**text2img.components)
+
+>>> # now you can use text2img(...) and img2img(...) just like the call methods of each respective pipeline
+```
+
+## VersatileDiffusionPipelineOutput
+[[autodoc]] pipelines.alt_diffusion.VersatileDiffusionPipelineOutput
+
+## VersatileDiffusionPipeline
+[[autodoc]] VersatileDiffusionPipeline
+	- __call__
+	- enable_attention_slicing
+	- disable_attention_slicing
+
+## VersatileDiffusionImg2ImgPipeline
+[[autodoc]] VersatileDiffusionImg2ImgPipeline
+	- __call__
+	- enable_attention_slicing
+	- disable_attention_slicing

scripts/convert_versatile_diffusion_to_diffusers.py

@@ -0,0 +1,791 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Conversion script for the Versatile Stable Diffusion checkpoints. """
+
+import argparse
+from argparse import Namespace
+
+import torch
+
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+    DPMSolverMultistepScheduler,
+    EulerAncestralDiscreteScheduler,
+    EulerDiscreteScheduler,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+    UNet2DConditionModel,
+    VersatileDiffusionPipeline,
+)
+from diffusers.pipelines.versatile_diffusion.modeling_text_unet import UNetFlatConditionModel
+from transformers import (
+    CLIPFeatureExtractor,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    CLIPVisionModelWithProjection,
+)
+
+
+SCHEDULER_CONFIG = Namespace(
+    **{
+        "beta_linear_start": 0.00085,
+        "beta_linear_end": 0.012,
+        "timesteps": 1000,
+        "scale_factor": 0.18215,
+    }
+)
+
+IMAGE_UNET_CONFIG = Namespace(
+    **{
+        "input_channels": 4,
+        "model_channels": 320,
+        "output_channels": 4,
+        "num_noattn_blocks": [2, 2, 2, 2],
+        "channel_mult": [1, 2, 4, 4],
+        "with_attn": [True, True, True, False],
+        "num_heads": 8,
+        "context_dim": 768,
+        "use_checkpoint": True,
+    }
+)
+
+TEXT_UNET_CONFIG = Namespace(
+    **{
+        "input_channels": 768,
+        "model_channels": 320,
+        "output_channels": 768,
+        "num_noattn_blocks": [2, 2, 2, 2],
+        "channel_mult": [1, 2, 4, 4],
+        "second_dim": [4, 4, 4, 4],
+        "with_attn": [True, True, True, False],
+        "num_heads": 8,
+        "context_dim": 768,
+        "use_checkpoint": True,
+    }
+)
+
+AUTOENCODER_CONFIG = Namespace(
+    **{
+        "double_z": True,
+        "z_channels": 4,
+        "resolution": 256,
+        "in_channels": 3,
+        "out_ch": 3,
+        "ch": 128,
+        "ch_mult": [1, 2, 4, 4],
+        "num_res_blocks": 2,
+        "attn_resolutions": [],
+        "dropout": 0.0,
+    }
+)
+
+
+def shave_segments(path, n_shave_prefix_segments=1):
+    """
+    Removes segments. Positive values shave the first segments, negative shave the last segments.
+    """
+    if n_shave_prefix_segments >= 0:
+        return ".".join(path.split(".")[n_shave_prefix_segments:])
+    else:
+        return ".".join(path.split(".")[:n_shave_prefix_segments])
+
+
+def renew_resnet_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside resnets to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item.replace("in_layers.0", "norm1")
+        new_item = new_item.replace("in_layers.2", "conv1")
+
+        new_item = new_item.replace("out_layers.0", "norm2")
+        new_item = new_item.replace("out_layers.3", "conv2")
+
+        new_item = new_item.replace("emb_layers.1", "time_emb_proj")
+        new_item = new_item.replace("skip_connection", "conv_shortcut")
+
+        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+        mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+def renew_vae_resnet_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside resnets to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item
+
+        new_item = new_item.replace("nin_shortcut", "conv_shortcut")
+        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+        mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+def renew_attention_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside attentions to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item
+
+        #         new_item = new_item.replace('norm.weight', 'group_norm.weight')
+        #         new_item = new_item.replace('norm.bias', 'group_norm.bias')
+
+        #         new_item = new_item.replace('proj_out.weight', 'proj_attn.weight')
+        #         new_item = new_item.replace('proj_out.bias', 'proj_attn.bias')
+
+        #         new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+        mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+def renew_vae_attention_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside attentions to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item
+
+        new_item = new_item.replace("norm.weight", "group_norm.weight")
+        new_item = new_item.replace("norm.bias", "group_norm.bias")
+
+        new_item = new_item.replace("q.weight", "query.weight")
+        new_item = new_item.replace("q.bias", "query.bias")
+
+        new_item = new_item.replace("k.weight", "key.weight")
+        new_item = new_item.replace("k.bias", "key.bias")
+
+        new_item = new_item.replace("v.weight", "value.weight")
+        new_item = new_item.replace("v.bias", "value.bias")
+
+        new_item = new_item.replace("proj_out.weight", "proj_attn.weight")
+        new_item = new_item.replace("proj_out.bias", "proj_attn.bias")
+
+        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+        mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+def assign_to_checkpoint(
+    paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None
+):
+    """
+    This does the final conversion step: take locally converted weights and apply a global renaming
+    to them. It splits attention layers, and takes into account additional replacements
+    that may arise.
+
+    Assigns the weights to the new checkpoint.
+    """
+    assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys."
+
+    # Splits the attention layers into three variables.
+    if attention_paths_to_split is not None:
+        for path, path_map in attention_paths_to_split.items():
+            old_tensor = old_checkpoint[path]
+            channels = old_tensor.shape[0] // 3
+
+            target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1)
+
+            num_heads = old_tensor.shape[0] // config["num_head_channels"] // 3
+
+            old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:])
+            query, key, value = old_tensor.split(channels // num_heads, dim=1)
+
+            checkpoint[path_map["query"]] = query.reshape(target_shape)
+            checkpoint[path_map["key"]] = key.reshape(target_shape)
+            checkpoint[path_map["value"]] = value.reshape(target_shape)
+
+    for path in paths:
+        new_path = path["new"]
+
+        # These have already been assigned
+        if attention_paths_to_split is not None and new_path in attention_paths_to_split:
+            continue
+
+        # Global renaming happens here
+        new_path = new_path.replace("middle_block.0", "mid_block.resnets.0")
+        new_path = new_path.replace("middle_block.1", "mid_block.attentions.0")
+        new_path = new_path.replace("middle_block.2", "mid_block.resnets.1")
+
+        if additional_replacements is not None:
+            for replacement in additional_replacements:
+                new_path = new_path.replace(replacement["old"], replacement["new"])
+
+        # proj_attn.weight has to be converted from conv 1D to linear
+        if "proj_attn.weight" in new_path:
+            checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0]
+        elif path["old"] in old_checkpoint:
+            checkpoint[new_path] = old_checkpoint[path["old"]]
+
+
+def conv_attn_to_linear(checkpoint):
+    keys = list(checkpoint.keys())
+    attn_keys = ["query.weight", "key.weight", "value.weight"]
+    for key in keys:
+        if ".".join(key.split(".")[-2:]) in attn_keys:
+            if checkpoint[key].ndim > 2:
+                checkpoint[key] = checkpoint[key][:, :, 0, 0]
+        elif "proj_attn.weight" in key:
+            if checkpoint[key].ndim > 2:
+                checkpoint[key] = checkpoint[key][:, :, 0]
+
+
+def create_image_unet_diffusers_config(unet_params):
+    """
+    Creates a config for the diffusers based on the config of the VD model.
+    """
+
+    block_out_channels = [unet_params.model_channels * mult for mult in unet_params.channel_mult]
+
+    down_block_types = []
+    resolution = 1
+    for i in range(len(block_out_channels)):
+        block_type = "CrossAttnDownBlock2D" if unet_params.with_attn[i] else "DownBlock2D"
+        down_block_types.append(block_type)
+        if i != len(block_out_channels) - 1:
+            resolution *= 2
+
+    up_block_types = []
+    for i in range(len(block_out_channels)):
+        block_type = "CrossAttnUpBlock2D" if unet_params.with_attn[-i - 1] else "UpBlock2D"
+        up_block_types.append(block_type)
+        resolution //= 2
+
+    if not all(n == unet_params.num_noattn_blocks[0] for n in unet_params.num_noattn_blocks):
+        raise ValueError("Not all num_res_blocks are equal, which is not supported in this script.")
+
+    config = dict(
+        sample_size=None,
+        in_channels=unet_params.input_channels,
+        out_channels=unet_params.output_channels,
+        down_block_types=tuple(down_block_types),
+        up_block_types=tuple(up_block_types),
+        block_out_channels=tuple(block_out_channels),
+        layers_per_block=unet_params.num_noattn_blocks[0],
+        cross_attention_dim=unet_params.context_dim,
+        attention_head_dim=unet_params.num_heads,
+    )
+
+    return config
+
+
+def create_text_unet_diffusers_config(unet_params):
+    """
+    Creates a config for the diffusers based on the config of the VD model.
+    """
+
+    block_out_channels = [unet_params.model_channels * mult for mult in unet_params.channel_mult]
+
+    down_block_types = []
+    resolution = 1
+    for i in range(len(block_out_channels)):
+        block_type = "CrossAttnDownBlockFlat" if unet_params.with_attn[i] else "DownBlockFlat"
+        down_block_types.append(block_type)
+        if i != len(block_out_channels) - 1:
+            resolution *= 2
+
+    up_block_types = []
+    for i in range(len(block_out_channels)):
+        block_type = "CrossAttnUpBlockFlat" if unet_params.with_attn[-i - 1] else "UpBlockFlat"
+        up_block_types.append(block_type)
+        resolution //= 2
+
+    if not all(n == unet_params.num_noattn_blocks[0] for n in unet_params.num_noattn_blocks):
+        raise ValueError("Not all num_res_blocks are equal, which is not supported in this script.")
+
+    config = dict(
+        sample_size=None,
+        in_channels=(unet_params.input_channels, 1, 1),
+        out_channels=(unet_params.output_channels, 1, 1),
+        down_block_types=tuple(down_block_types),
+        up_block_types=tuple(up_block_types),
+        block_out_channels=tuple(block_out_channels),
+        layers_per_block=unet_params.num_noattn_blocks[0],
+        cross_attention_dim=unet_params.context_dim,
+        attention_head_dim=unet_params.num_heads,
+    )
+
+    return config
+
+
+def create_vae_diffusers_config(vae_params):
+    """
+    Creates a config for the diffusers based on the config of the VD model.
+    """
+
+    block_out_channels = [vae_params.ch * mult for mult in vae_params.ch_mult]
+    down_block_types = ["DownEncoderBlock2D"] * len(block_out_channels)
+    up_block_types = ["UpDecoderBlock2D"] * len(block_out_channels)
+
+    config = dict(
+        sample_size=vae_params.resolution,
+        in_channels=vae_params.in_channels,
+        out_channels=vae_params.out_ch,
+        down_block_types=tuple(down_block_types),
+        up_block_types=tuple(up_block_types),
+        block_out_channels=tuple(block_out_channels),
+        latent_channels=vae_params.z_channels,
+        layers_per_block=vae_params.num_res_blocks,
+    )
+    return config
+
+
+def create_diffusers_scheduler(original_config):
+    schedular = DDIMScheduler(
+        num_train_timesteps=original_config.model.params.timesteps,
+        beta_start=original_config.model.params.linear_start,
+        beta_end=original_config.model.params.linear_end,
+        beta_schedule="scaled_linear",
+    )
+    return schedular
+
+
+def convert_vd_unet_checkpoint(checkpoint, config, unet_key, extract_ema=False):
+    """
+    Takes a state dict and a config, and returns a converted checkpoint.
+    """
+
+    # extract state_dict for UNet
+    unet_state_dict = {}
+    keys = list(checkpoint.keys())
+
+    # at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA
+    if sum(k.startswith("model_ema") for k in keys) > 100:
+        print("Checkpoint has both EMA and non-EMA weights.")
+        if extract_ema:
+            print(
+                "In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA"
+                " weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag."
+            )
+            for key in keys:
+                if key.startswith("model.diffusion_model"):
+                    flat_ema_key = "model_ema." + "".join(key.split(".")[1:])
+                    unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(flat_ema_key)
+        else:
+            print(
+                "In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA"
+                " weights (usually better for inference), please make sure to add the `--extract_ema` flag."
+            )
+
+    for key in keys:
+        if key.startswith(unet_key):
+            unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(key)
+
+    new_checkpoint = {}
+
+    new_checkpoint["time_embedding.linear_1.weight"] = checkpoint["model.diffusion_model.time_embed.0.weight"]
+    new_checkpoint["time_embedding.linear_1.bias"] = checkpoint["model.diffusion_model.time_embed.0.bias"]
+    new_checkpoint["time_embedding.linear_2.weight"] = checkpoint["model.diffusion_model.time_embed.2.weight"]
+    new_checkpoint["time_embedding.linear_2.bias"] = checkpoint["model.diffusion_model.time_embed.2.bias"]
+
+    new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"]
+    new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"]
+
+    new_checkpoint["conv_norm_out.weight"] = unet_state_dict["out.0.weight"]
+    new_checkpoint["conv_norm_out.bias"] = unet_state_dict["out.0.bias"]
+    new_checkpoint["conv_out.weight"] = unet_state_dict["out.2.weight"]
+    new_checkpoint["conv_out.bias"] = unet_state_dict["out.2.bias"]
+
+    # Retrieves the keys for the input blocks only
+    num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer})
+    input_blocks = {
+        layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}" in key]
+        for layer_id in range(num_input_blocks)
+    }
+
+    # Retrieves the keys for the middle blocks only
+    num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer})
+    middle_blocks = {
+        layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}" in key]
+        for layer_id in range(num_middle_blocks)
+    }
+
+    # Retrieves the keys for the output blocks only
+    num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer})
+    output_blocks = {
+        layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}" in key]
+        for layer_id in range(num_output_blocks)
+    }
+
+    for i in range(1, num_input_blocks):
+        block_id = (i - 1) // (config["layers_per_block"] + 1)
+        layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1)
+
+        resnets = [
+            key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key
+        ]
+        attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key]
+
+        if f"input_blocks.{i}.0.op.weight" in unet_state_dict:
+            new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop(
+                f"input_blocks.{i}.0.op.weight"
+            )
+            new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop(
+                f"input_blocks.{i}.0.op.bias"
+            )
+        elif f"input_blocks.{i}.0.weight" in unet_state_dict:
+            # text_unet uses linear layers in place of downsamplers
+            shape = unet_state_dict[f"input_blocks.{i}.0.weight"].shape
+            if shape[0] != shape[1]:
+                continue
+            new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.weight"] = unet_state_dict.pop(
+                f"input_blocks.{i}.0.weight"
+            )
+            new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.bias"] = unet_state_dict.pop(
+                f"input_blocks.{i}.0.bias"
+            )
+
+        paths = renew_resnet_paths(resnets)
+        meta_path = {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"}
+        assign_to_checkpoint(
+            paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+        )
+
+        if len(attentions):
+            paths = renew_attention_paths(attentions)
+            meta_path = {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"}
+            assign_to_checkpoint(
+                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+            )
+
+    resnet_0 = middle_blocks[0]
+    attentions = middle_blocks[1]
+    resnet_1 = middle_blocks[2]
+
+    resnet_0_paths = renew_resnet_paths(resnet_0)
+    assign_to_checkpoint(resnet_0_paths, new_checkpoint, unet_state_dict, config=config)
+
+    resnet_1_paths = renew_resnet_paths(resnet_1)
+    assign_to_checkpoint(resnet_1_paths, new_checkpoint, unet_state_dict, config=config)
+
+    attentions_paths = renew_attention_paths(attentions)
+    meta_path = {"old": "middle_block.1", "new": "mid_block.attentions.0"}
+    assign_to_checkpoint(
+        attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+    )
+
+    for i in range(num_output_blocks):
+        block_id = i // (config["layers_per_block"] + 1)
+        layer_in_block_id = i % (config["layers_per_block"] + 1)
+        output_block_layers = [shave_segments(name, 2) for name in output_blocks[i]]
+        output_block_list = {}
+
+        for layer in output_block_layers:
+            layer_id, layer_name = layer.split(".")[0], shave_segments(layer, 1)
+            if layer_id in output_block_list:
+                output_block_list[layer_id].append(layer_name)
+            else:
+                output_block_list[layer_id] = [layer_name]
+
+        if len(output_block_list) > 1:
+            resnets = [key for key in output_blocks[i] if f"output_blocks.{i}.0" in key]
+            attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.1" in key]
+
+            paths = renew_resnet_paths(resnets)
+
+            meta_path = {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"}
+            assign_to_checkpoint(
+                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+            )
+
+            if ["conv.weight", "conv.bias"] in output_block_list.values():
+                index = list(output_block_list.values()).index(["conv.weight", "conv.bias"])
+                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[
+                    f"output_blocks.{i}.{index}.conv.weight"
+                ]
+                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[
+                    f"output_blocks.{i}.{index}.conv.bias"
+                ]
+                # Clear attentions as they have been attributed above.
+                if len(attentions) == 2:
+                    attentions = []
+            elif f"output_blocks.{i}.1.weight" in unet_state_dict:
+                # text_unet uses linear layers in place of upsamplers
+                shape = unet_state_dict[f"output_blocks.{i}.1.weight"].shape
+                if shape[0] != shape[1]:
+                    continue
+                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.weight"] = unet_state_dict.pop(
+                    f"output_blocks.{i}.1.weight"
+                )
+                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.bias"] = unet_state_dict.pop(
+                    f"output_blocks.{i}.1.bias"
+                )
+                # Clear attentions as they have been attributed above.
+                if len(attentions) == 2:
+                    attentions = []
+            elif f"output_blocks.{i}.2.weight" in unet_state_dict:
+                # text_unet uses linear layers in place of upsamplers
+                shape = unet_state_dict[f"output_blocks.{i}.2.weight"].shape
+                if shape[0] != shape[1]:
+                    continue
+                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.weight"] = unet_state_dict.pop(
+                    f"output_blocks.{i}.2.weight"
+                )
+                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.bias"] = unet_state_dict.pop(
+                    f"output_blocks.{i}.2.bias"
+                )
+
+            if len(attentions):
+                paths = renew_attention_paths(attentions)
+                meta_path = {
+                    "old": f"output_blocks.{i}.1",
+                    "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}",
+                }
+                assign_to_checkpoint(
+                    paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+                )
+        else:
+            resnet_0_paths = renew_resnet_paths(output_block_layers, n_shave_prefix_segments=1)
+            for path in resnet_0_paths:
+                old_path = ".".join(["output_blocks", str(i), path["old"]])
+                new_path = ".".join(["up_blocks", str(block_id), "resnets", str(layer_in_block_id), path["new"]])
+
+                new_checkpoint[new_path] = unet_state_dict[old_path]
+
+    return new_checkpoint
+
+
+def convert_vd_vae_checkpoint(checkpoint, config):
+    # extract state dict for VAE
+    vae_state_dict = {}
+    keys = list(checkpoint.keys())
+    for key in keys:
+        vae_state_dict[key] = checkpoint.get(key)
+
+    new_checkpoint = {}
+
+    new_checkpoint["encoder.conv_in.weight"] = vae_state_dict["encoder.conv_in.weight"]
+    new_checkpoint["encoder.conv_in.bias"] = vae_state_dict["encoder.conv_in.bias"]
+    new_checkpoint["encoder.conv_out.weight"] = vae_state_dict["encoder.conv_out.weight"]
+    new_checkpoint["encoder.conv_out.bias"] = vae_state_dict["encoder.conv_out.bias"]
+    new_checkpoint["encoder.conv_norm_out.weight"] = vae_state_dict["encoder.norm_out.weight"]
+    new_checkpoint["encoder.conv_norm_out.bias"] = vae_state_dict["encoder.norm_out.bias"]
+
+    new_checkpoint["decoder.conv_in.weight"] = vae_state_dict["decoder.conv_in.weight"]
+    new_checkpoint["decoder.conv_in.bias"] = vae_state_dict["decoder.conv_in.bias"]
+    new_checkpoint["decoder.conv_out.weight"] = vae_state_dict["decoder.conv_out.weight"]
+    new_checkpoint["decoder.conv_out.bias"] = vae_state_dict["decoder.conv_out.bias"]
+    new_checkpoint["decoder.conv_norm_out.weight"] = vae_state_dict["decoder.norm_out.weight"]
+    new_checkpoint["decoder.conv_norm_out.bias"] = vae_state_dict["decoder.norm_out.bias"]
+
+    new_checkpoint["quant_conv.weight"] = vae_state_dict["quant_conv.weight"]
+    new_checkpoint["quant_conv.bias"] = vae_state_dict["quant_conv.bias"]
+    new_checkpoint["post_quant_conv.weight"] = vae_state_dict["post_quant_conv.weight"]
+    new_checkpoint["post_quant_conv.bias"] = vae_state_dict["post_quant_conv.bias"]
+
+    # Retrieves the keys for the encoder down blocks only
+    num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "encoder.down" in layer})
+    down_blocks = {
+        layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)
+    }
+
+    # Retrieves the keys for the decoder up blocks only
+    num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "decoder.up" in layer})
+    up_blocks = {
+        layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)
+    }
+
+    for i in range(num_down_blocks):
+        resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key]
+
+        if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict:
+            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop(
+                f"encoder.down.{i}.downsample.conv.weight"
+            )
+            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop(
+                f"encoder.down.{i}.downsample.conv.bias"
+            )
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"}
+        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+    mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key]
+    num_mid_res_blocks = 2
+    for i in range(1, num_mid_res_blocks + 1):
+        resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key]
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
+        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+    mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key]
+    paths = renew_vae_attention_paths(mid_attentions)
+    meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+    assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+    conv_attn_to_linear(new_checkpoint)
+
+    for i in range(num_up_blocks):
+        block_id = num_up_blocks - 1 - i
+        resnets = [
+            key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key
+        ]
+
+        if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict:
+            new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[
+                f"decoder.up.{block_id}.upsample.conv.weight"
+            ]
+            new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[
+                f"decoder.up.{block_id}.upsample.conv.bias"
+            ]
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"}
+        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+    mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key]
+    num_mid_res_blocks = 2
+    for i in range(1, num_mid_res_blocks + 1):
+        resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key]
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
+        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+    mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key]
+    paths = renew_vae_attention_paths(mid_attentions)
+    meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+    assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+    conv_attn_to_linear(new_checkpoint)
+    return new_checkpoint
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--unet_checkpoint_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--vae_checkpoint_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--optimus_checkpoint_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--scheduler_type",
+        default="pndm",
+        type=str,
+        help="Type of scheduler to use. Should be one of ['pndm', 'lms', 'ddim', 'euler', 'euler-ancest', 'dpm']",
+    )
+    parser.add_argument(
+        "--extract_ema",
+        action="store_true",
+        help=(
+            "Only relevant for checkpoints that have both EMA and non-EMA weights. Whether to extract the EMA weights"
+            " or not. Defaults to `False`. Add `--extract_ema` to extract the EMA weights. EMA weights usually yield"
+            " higher quality images for inference. Non-EMA weights are usually better to continue fine-tuning."
+        ),
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output model.")
+
+    args = parser.parse_args()
+
+    scheduler_config = SCHEDULER_CONFIG
+
+    num_train_timesteps = scheduler_config.timesteps
+    beta_start = scheduler_config.beta_linear_start
+    beta_end = scheduler_config.beta_linear_end
+    if args.scheduler_type == "pndm":
+        scheduler = PNDMScheduler(
+            beta_end=beta_end,
+            beta_schedule="scaled_linear",
+            beta_start=beta_start,
+            num_train_timesteps=num_train_timesteps,
+            skip_prk_steps=True,
+            steps_offset=1,
+        )
+    elif args.scheduler_type == "lms":
+        scheduler = LMSDiscreteScheduler(beta_start=beta_start, beta_end=beta_end, beta_schedule="scaled_linear")
+    elif args.scheduler_type == "euler":
+        scheduler = EulerDiscreteScheduler(beta_start=beta_start, beta_end=beta_end, beta_schedule="scaled_linear")
+    elif args.scheduler_type == "euler-ancestral":
+        scheduler = EulerAncestralDiscreteScheduler(
+            beta_start=beta_start, beta_end=beta_end, beta_schedule="scaled_linear"
+        )
+    elif args.scheduler_type == "dpm":
+        scheduler = DPMSolverMultistepScheduler(
+            beta_start=beta_start, beta_end=beta_end, beta_schedule="scaled_linear"
+        )
+    elif args.scheduler_type == "ddim":
+        scheduler = DDIMScheduler(
+            beta_start=beta_start,
+            beta_end=beta_end,
+            beta_schedule="scaled_linear",
+            clip_sample=False,
+            set_alpha_to_one=False,
+            steps_offset=1,
+        )
+    else:
+        raise ValueError(f"Scheduler of type {args.scheduler_type} doesn't exist!")
+
+    # Convert the UNet2DConditionModel models.
+    if args.unet_checkpoint_path is not None:
+        # image UNet
+        image_unet_config = create_image_unet_diffusers_config(IMAGE_UNET_CONFIG)
+        checkpoint = torch.load(args.unet_checkpoint_path)
+        converted_image_unet_checkpoint = convert_vd_unet_checkpoint(
+            checkpoint, image_unet_config, unet_key="model.diffusion_model.unet_image.", extract_ema=args.extract_ema
+        )
+        image_unet = UNet2DConditionModel(**image_unet_config)
+        image_unet.load_state_dict(converted_image_unet_checkpoint)
+
+        # text UNet
+        text_unet_config = create_text_unet_diffusers_config(TEXT_UNET_CONFIG)
+        converted_text_unet_checkpoint = convert_vd_unet_checkpoint(
+            checkpoint, text_unet_config, unet_key="model.diffusion_model.unet_text.", extract_ema=args.extract_ema
+        )
+        text_unet = UNetFlatConditionModel(**text_unet_config)
+        text_unet.load_state_dict(converted_text_unet_checkpoint)
+
+    # Convert the VAE model.
+    if args.vae_checkpoint_path is not None:
+        vae_config = create_vae_diffusers_config(AUTOENCODER_CONFIG)
+        checkpoint = torch.load(args.vae_checkpoint_path)
+        converted_vae_checkpoint = convert_vd_vae_checkpoint(checkpoint, vae_config)
+
+        vae = AutoencoderKL(**vae_config)
+        vae.load_state_dict(converted_vae_checkpoint)
+
+    tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
+    image_feature_extractor = CLIPFeatureExtractor.from_pretrained("openai/clip-vit-large-patch14")
+    text_encoder = CLIPTextModelWithProjection.from_pretrained("openai/clip-vit-large-patch14")
+    image_encoder = CLIPVisionModelWithProjection.from_pretrained("openai/clip-vit-large-patch14")
+
+    pipe = VersatileDiffusionPipeline(
+        scheduler=scheduler,
+        tokenizer=tokenizer,
+        image_feature_extractor=image_feature_extractor,
+        text_encoder=text_encoder,
+        image_encoder=image_encoder,
+        image_unet=image_unet,
+        text_unet=text_unet,
+        vae=vae,
+    )
+    pipe.save_pretrained(args.dump_path)

src/diffusers/__init__.py

@@ -73,6 +73,11 @@ if is_torch_available() and is_transformers_available():
         StableDiffusionInpaintPipeline,
         StableDiffusionInpaintPipelineLegacy,
         StableDiffusionPipeline,
+        VersatileDiffusionDualGuidedPipeline,
+        VersatileDiffusionImageToTextPipeline,
+        VersatileDiffusionImageVariationPipeline,
+        VersatileDiffusionPipeline,
+        VersatileDiffusionTextToImagePipeline,
         VQDiffusionPipeline,
     )
 else:

src/diffusers/models/attention.py

@@ -22,7 +22,7 @@ from torch import nn
 from ..configuration_utils import ConfigMixin, register_to_config
 from ..modeling_utils import ModelMixin
 from ..models.embeddings import ImagePositionalEmbeddings
-from ..utils import BaseOutput
+from ..utils import CONFIG_NAME, BaseOutput
 from ..utils.import_utils import is_xformers_available
 
 
@@ -666,3 +666,120 @@ class AdaLayerNorm(nn.Module):
         scale, shift = torch.chunk(emb, 2)
         x = self.norm(x) * (1 + scale) + shift
         return x
+
+
+class DualTransformer2DModel(nn.Module):
+    """
+    Dual transformer wrapper that combines two `Transformer2DModel`s for mixed inference.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            Pass if the input is continuous. The number of channels in the input and output.
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of context dimensions to use.
+        sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
+            Note that this is fixed at training time as it is used for learning a number of position embeddings. See
+            `ImagePositionalEmbeddings`.
+        num_vector_embeds (`int`, *optional*):
+            Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
+            The number of diffusion steps used during training. Note that this is fixed at training time as it is used
+            to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
+            up to but not more than steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the TransformerBlocks' attention should contain a bias parameter.
+    """
+
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+    ):
+        super().__init__()
+        self.transformers = nn.ModuleList(
+            [
+                Transformer2DModel(
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    in_channels=in_channels,
+                    num_layers=num_layers,
+                    dropout=dropout,
+                    norm_num_groups=norm_num_groups,
+                    cross_attention_dim=cross_attention_dim,
+                    attention_bias=attention_bias,
+                    sample_size=sample_size,
+                    num_vector_embeds=num_vector_embeds,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                )
+                for _ in range(2)
+            ]
+        )
+
+        # The ratio of transformer1 to transformer2's output states to be combined during inference
+        self.mix_ratio = 0.5
+
+        # The shape of `encoder_hidden_states` is expected to be
+        # `(batch_size, num_condition_tokens[0]+num_condition_tokens[1], num_features)`
+        self.num_condition_tokens = (77, 257)
+
+    def forward(self, hidden_states, encoder_hidden_states, timestep=None, return_dict: bool = True):
+        """
+        Args:
+            hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
+                When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input
+                hidden_states
+            encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, context dim)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.long`, *optional*):
+                Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.attention.Transformer2DModelOutput`] or `tuple`: [`~models.attention.Transformer2DModelOutput`]
+            if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample
+            tensor.
+        """
+        input_states = hidden_states
+
+        encoded_states = []
+        tokens_start = 0
+        for i in range(2):
+            # for each of the two transformers, pass the corresponding condition tokens
+            condition_state = encoder_hidden_states[:, tokens_start : tokens_start + self.num_condition_tokens[i]]
+            encoded_state = self.transformers[i](input_states, condition_state, timestep, return_dict)[0]
+            encoded_states.append(encoded_state - input_states)
+            tokens_start += self.num_condition_tokens[i]
+
+        output_states = encoded_states[0] * self.mix_ratio + encoded_states[1] * (1 - self.mix_ratio)
+        output_states = output_states + input_states
+
+        if not return_dict:
+            return (output_states,)
+
+        return Transformer2DModelOutput(sample=output_states)
+
+    def _set_attention_slice(self, slice_size):
+        for transformer in self.transformers:
+            transformer._set_attention_slice(slice_size)
+
+    def _set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        for transformer in self.transformers:
+            transformer._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)

src/diffusers/models/unet_2d.py

@@ -175,7 +175,7 @@ class UNet2DModel(ModelMixin, ConfigMixin):
         num_groups_out = norm_num_groups if norm_num_groups is not None else min(block_out_channels[0] // 4, 32)
         self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=num_groups_out, eps=norm_eps)
         self.conv_act = nn.SiLU()
-        self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
+        self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
 
     def forward(
         self,

src/diffusers/models/unet_2d_blocks.py

@@ -15,7 +15,7 @@ import numpy as np
 import torch
 from torch import nn
 
-from .attention import AttentionBlock, Transformer2DModel
+from .attention import AttentionBlock, Transformer2DModel, DualTransformer2DModel
 from .resnet import Downsample2D, FirDownsample2D, FirUpsample2D, ResnetBlock2D, Upsample2D
 
 
@@ -32,6 +32,7 @@ def get_down_block(
     resnet_groups=None,
     cross_attention_dim=None,
     downsample_padding=None,
+    dual_cross_attention=False,
 ):
     down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
     if down_block_type == "DownBlock2D":
@@ -74,6 +75,7 @@ def get_down_block(
             downsample_padding=downsample_padding,
             cross_attention_dim=cross_attention_dim,
             attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
         )
     elif down_block_type == "SkipDownBlock2D":
         return SkipDownBlock2D(
@@ -137,6 +139,7 @@ def get_up_block(
     attn_num_head_channels,
     resnet_groups=None,
     cross_attention_dim=None,
+    dual_cross_attention=False,
 ):
     up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
     if up_block_type == "UpBlock2D":
@@ -322,6 +325,7 @@ class UNetMidBlock2DCrossAttn(nn.Module):
         attention_type="default",
         output_scale_factor=1.0,
         cross_attention_dim=1280,
+        dual_cross_attention=False,
         **kwargs,
     ):
         super().__init__()
@@ -505,6 +509,7 @@ class CrossAttnDownBlock2D(nn.Module):
         output_scale_factor=1.0,
         downsample_padding=1,
         add_downsample=True,
+        dual_cross_attention=False,
     ):
         super().__init__()
         resnets = []
@@ -529,16 +534,28 @@ class CrossAttnDownBlock2D(nn.Module):
                     pre_norm=resnet_pre_norm,
                 )
             )
-            attentions.append(
-                Transformer2DModel(
-                    attn_num_head_channels,
-                    out_channels // attn_num_head_channels,
-                    in_channels=out_channels,
-                    num_layers=1,
-                    cross_attention_dim=cross_attention_dim,
-                    norm_num_groups=resnet_groups,
+            if dual_cross_attention is False:
+                attentions.append(
+                    Transformer2DModel(
+                        attn_num_head_channels,
+                        out_channels // attn_num_head_channels,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        attn_num_head_channels,
+                        out_channels // attn_num_head_channels,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
                 )
-            )
         self.attentions = nn.ModuleList(attentions)
         self.resnets = nn.ModuleList(resnets)
 

src/diffusers/models/unet_2d_condition.py

@@ -106,6 +106,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin):
         norm_eps: float = 1e-5,
         cross_attention_dim: int = 1280,
         attention_head_dim: int = 8,
+        dual_cross_attention: bool = False,
     ):
         super().__init__()
 
@@ -145,6 +146,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin):
                 cross_attention_dim=cross_attention_dim,
                 attn_num_head_channels=attention_head_dim,
                 downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
             )
             self.down_blocks.append(down_block)
 
@@ -159,6 +161,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin):
             cross_attention_dim=cross_attention_dim,
             attn_num_head_channels=attention_head_dim,
             resnet_groups=norm_num_groups,
+            dual_cross_attention=dual_cross_attention,
         )
 
         # count how many layers upsample the images
@@ -194,6 +197,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin):
                 resnet_groups=norm_num_groups,
                 cross_attention_dim=cross_attention_dim,
                 attn_num_head_channels=attention_head_dim,
+                dual_cross_attention=dual_cross_attention,
             )
             self.up_blocks.append(up_block)
             prev_output_channel = output_channel
@@ -201,7 +205,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin):
         # out
         self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps)
         self.conv_act = nn.SiLU()
-        self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
+        self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
 
     def set_attention_slice(self, slice_size):
         if slice_size is not None and self.config.attention_head_dim % slice_size != 0:

src/diffusers/pipelines/__init__.py

@@ -24,6 +24,13 @@ if is_torch_available() and is_transformers_available():
         StableDiffusionInpaintPipelineLegacy,
         StableDiffusionPipeline,
     )
+    from .versatile_diffusion import (
+        VersatileDiffusionDualGuidedPipeline,
+        VersatileDiffusionImageToTextPipeline,
+        VersatileDiffusionImageVariationPipeline,
+        VersatileDiffusionPipeline,
+        VersatileDiffusionTextToImagePipeline,
+    )
     from .vq_diffusion import VQDiffusionPipeline
 
 if is_transformers_available() and is_onnx_available():

src/diffusers/pipelines/versatile_diffusion/__init__.py

@@ -0,0 +1,11 @@
+from ...utils import is_torch_available, is_transformers_available
+
+
+if is_transformers_available() and is_torch_available():
+    from .modeling_gpt2_optimus import GPT2OptimusForLatentConnector
+    from .modeling_text_unet import UNetFlatConditionModel
+    from .pipeline_versatile_diffusion import VersatileDiffusionPipeline
+    from .pipeline_versatile_diffusion_dual_guided import VersatileDiffusionDualGuidedPipeline
+    from .pipeline_versatile_diffusion_image_to_text import VersatileDiffusionImageToTextPipeline
+    from .pipeline_versatile_diffusion_image_variation import VersatileDiffusionImageVariationPipeline
+    from .pipeline_versatile_diffusion_text_to_image import VersatileDiffusionTextToImagePipeline

src/diffusers/pipelines/versatile_diffusion/modeling_gpt2_optimus.py

@@ -0,0 +1,345 @@
+import math
+
+import torch
+from torch import nn
+
+from transformers.modeling_outputs import CausalLMOutputWithCrossAttentions
+from transformers.models.gpt2.modeling_gpt2 import GPT2MLP, GPT2PreTrainedModel
+from transformers.pytorch_utils import Conv1D
+
+
+class GPT2OptimusAttention(nn.Module):
+    def __init__(self, nx, n_ctx, config, scale=False):
+        super().__init__()
+        self.output_attentions = config.output_attentions
+
+        n_state = nx  # in Attention: n_state=768 (nx=n_embd)
+        # [switch nx => n_state from Block to Attention to keep identical to TF implem]
+        assert n_state % config.n_head == 0
+        self.register_buffer("bias", torch.tril(torch.ones(n_ctx, n_ctx)).view(1, 1, n_ctx, n_ctx))
+        self.n_head = config.n_head
+        self.split_size = n_state
+        self.scale = scale
+
+        self.c_attn = Conv1D(n_state * 3, nx)
+        self.c_proj = Conv1D(n_state, nx)
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.pruned_heads = set()
+
+    def _attn(self, q, k, v, attention_mask=None, head_mask=None):
+        w = torch.matmul(q, k)
+        if self.scale:
+            w = w / math.sqrt(v.size(-1))
+        nd, ns = w.size(-2), w.size(-1)
+        b = self.bias[:, :, ns - nd : ns, :ns]
+        w = w * b - 1e4 * (1 - b)
+
+        if attention_mask is not None:
+            # Apply the attention mask
+            w = w + attention_mask
+
+        w = nn.Softmax(dim=-1)(w)
+        w = self.attn_dropout(w)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            w = w * head_mask
+
+        outputs = [torch.matmul(w, v)]
+        if self.output_attentions:
+            outputs.append(w)
+        return outputs
+
+    def merge_heads(self, x):
+        x = x.permute(0, 2, 1, 3).contiguous()
+        new_x_shape = x.size()[:-2] + (x.size(-2) * x.size(-1),)
+        return x.view(*new_x_shape)  # in Tensorflow implem: fct merge_states
+
+    def split_heads(self, x, k=False):
+        new_x_shape = x.size()[:-1] + (self.n_head, x.size(-1) // self.n_head)
+        x = x.view(*new_x_shape)  # in Tensorflow implem: fct split_states
+        if k:
+            return x.permute(0, 2, 3, 1)  # (batch, head, head_features, seq_length)
+        else:
+            return x.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)
+
+    def forward(self, x, layer_past=None, attention_mask=None, head_mask=None):
+        x = self.c_attn(x)
+        query, key, value = x.split(self.split_size, dim=2)
+        query = self.split_heads(query)
+        key = self.split_heads(key, k=True)
+        value = self.split_heads(value)
+
+        if layer_past is not None:
+            past_key, past_value = layer_past[0], layer_past[1]  # transpose back cf below
+
+            past_key = self.split_heads(past_key, k=True)
+            past_value = self.split_heads(past_value)
+            # pdb.set_trace()
+            key = torch.cat((past_key, key), dim=-1)
+            value = torch.cat((past_value, value), dim=-2)
+        present = torch.stack((key.transpose(-2, -1), value))  # transpose to have same shapes for stacking
+
+        attn_outputs = self._attn(query, key, value, attention_mask, head_mask)
+        a = attn_outputs[0]
+
+        a = self.merge_heads(a)
+        a = self.c_proj(a)
+        a = self.resid_dropout(a)
+
+        outputs = [a, present] + attn_outputs[1:]
+        return outputs  # a, present, (attentions)
+
+
+class GPT2OptimusBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        nx = config.n_embd
+        self.ln_1 = nn.LayerNorm(nx, eps=config.layer_norm_epsilon)
+        self.attn = GPT2OptimusAttention(nx, config.n_ctx, config, scale=True)
+        self.ln_2 = nn.LayerNorm(nx, eps=config.layer_norm_epsilon)
+        self.mlp = GPT2MLP(4 * nx, config)
+
+    def forward(self, x, layer_past=None, attention_mask=None, head_mask=None):
+        output_attn = self.attn(
+            self.ln_1(x), layer_past=layer_past, attention_mask=attention_mask, head_mask=head_mask
+        )
+        a = output_attn[0]  # output_attn: a, present, (attentions)
+
+        x = x + a
+        m = self.mlp(self.ln_2(x))
+        x = x + m
+
+        outputs = [x] + output_attn[1:]
+        return outputs  # x, present, (attentions)
+
+
+class GPT2OptimusModel(GPT2PreTrainedModel):
+    def __init__(self, config, latent_as_gpt_emb, latent_as_gpt_memory, latent_size):
+        super().__init__(config)
+        self.latent_as_gpt_emb = latent_as_gpt_emb
+        self.latent_as_gpt_memory = latent_as_gpt_memory
+        self.latent_size = latent_size
+        self.output_hidden_states = config.output_hidden_states
+        self.output_attentions = config.output_attentions
+
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+        self.wpe = nn.Embedding(config.n_positions, config.n_embd)
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList([GPT2OptimusBlock(config) for i in range(config.n_layer)])
+        self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+
+        self.linear = nn.Linear(
+            self.latent_size, config.hidden_size * config.n_layer, bias=False
+        )  # different latent vector for each layer
+        self.linear_emb = nn.Linear(
+            self.latent_size, config.hidden_size, bias=False
+        )  # share the same latent vector as the embeddings
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids,
+        past=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+    ):
+        if past is None:
+            past_length = 0
+            past = [None] * len(self.h)
+        else:
+            if self.latent_as_gpt_emb:
+                past_emb = self.linear_emb(past)  # used as embeddings to add on other three embeddings
+
+            if self.latent_as_gpt_memory:
+                past = self.linear(past)
+
+                # different latent vectors for each layer
+                past_split = torch.split(past.unsqueeze(1), self.config.hidden_size, dim=2)
+                past = list(zip(past_split, past_split))
+                past_length = 1  # past[0][0].size(-2)
+            else:
+                past_length = 0
+                past = [None] * len(self.h)
+
+        if position_ids is None:
+            position_ids = torch.arange(
+                past_length, input_ids.size(-1) + past_length, dtype=torch.long, device=input_ids.device
+            )
+            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
+
+        # Attention mask.
+        if attention_mask is not None:
+            # We create a 3D attention mask from a 2D tensor mask.
+            # Sizes are [batch_size, 1, 1, to_seq_length]
+            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            # this attention mask is more simple than the triangular masking of causal attention
+            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+            attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+
+            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+            # masked positions, this operation will create a tensor which is 0.0 for
+            # positions we want to attend and -10000.0 for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            attention_mask = attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # head_mask has shape n_layer x batch x n_heads x N x N
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.n_layer, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = (
+                    head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)
+                )  # We can specify head_mask for each layer
+            head_mask = head_mask.to(
+                dtype=next(self.parameters()).dtype
+            )  # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.n_layer
+
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_ids.size(-1))
+        position_ids = position_ids.view(-1, position_ids.size(-1))
+
+        inputs_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1))
+            token_type_embeds = self.wte(token_type_ids)
+        else:
+            token_type_embeds = 0
+
+        hidden_states = inputs_embeds + position_embeds + token_type_embeds
+        if self.latent_as_gpt_emb:
+            hidden_states = hidden_states + past_emb.unsqueeze(1)
+
+        hidden_states = self.drop(hidden_states)
+
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        presents = ()
+        all_attentions = []
+        all_hidden_states = ()
+        for i, (block, layer_past) in enumerate(zip(self.h, past)):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)
+
+            outputs = block(
+                hidden_states, layer_past=layer_past, attention_mask=attention_mask, head_mask=head_mask[i]
+            )
+
+            hidden_states, present = outputs[:2]
+            presents = presents + (present,)
+
+            if self.output_attentions:
+                all_attentions.append(outputs[2])
+
+        hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(*output_shape)
+        # Add last hidden state
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states, presents)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            # let the number of heads free (-1) so we can extract attention even after head pruning
+            attention_output_shape = input_shape[:-1] + (-1,) + all_attentions[0].shape[-2:]
+            all_attentions = tuple(t.view(*attention_output_shape) for t in all_attentions)
+            outputs = outputs + (all_attentions,)
+
+        return outputs  # last hidden state, presents, (all hidden_states), (attentions)
+
+
+class GPT2OptimusForLatentConnector(GPT2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.latent_as_gpt_emb = True
+        self.latent_as_gpt_memory = True
+        self.latent_size = getattr(config, "latent_size", 32)
+        self.transformer = GPT2OptimusModel(
+            config,
+            latent_as_gpt_emb=self.latent_as_gpt_emb,
+            latent_as_gpt_memory=self.latent_as_gpt_memory,
+            latent_size=self.latent_size,
+        )
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        self.init_weights()
+        self.tie_weights()
+
+        # Initialize weights and apply final processing
+        self.post_init()
+        self.tie_weights()
+
+    def _tie_or_clone_weights(self, first_module, second_module):
+        """Tie or clone module weights depending of weither we are using TorchScript or not"""
+        if self.config.torchscript:
+            first_module.weight = nn.Parameter(second_module.weight.clone())
+        else:
+            first_module.weight = second_module.weight
+
+        if hasattr(first_module, "bias") and first_module.bias is not None:
+            first_module.bias.data = torch.nn.functional.pad(
+                first_module.bias.data,
+                (0, first_module.weight.shape[0] - first_module.bias.shape[0]),
+                "constant",
+                0,
+            )
+
+    def tie_weights(self):
+        """Make sure we are sharing the input and output embeddings.
+        Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        self._tie_or_clone_weights(self.lm_head, self.transformer.wte)
+
+    def forward(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=True,
+    ):
+        transformer_outputs = self.transformer(
+            input_ids,
+            past=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+        )
+        hidden_states = transformer_outputs[0]
+
+        lm_logits = self.lm_head(hidden_states)
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=None,
+            logits=lm_logits,
+            past_key_values=past_key_values,
+            hidden_states=None,
+            attentions=None,
+            cross_attentions=None,
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past, **kwargs):
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+        }

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion.py

@@ -0,0 +1,180 @@
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+
+import PIL.Image
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer, CLIPVisionModel
+
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...pipeline_utils import DiffusionPipeline
+from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ...utils import logging
+from .pipeline_versatile_diffusion_image_variation import VersatileDiffusionImageVariationPipeline
+from .pipeline_versatile_diffusion_text_to_image import VersatileDiffusionTextToImagePipeline
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class VersatileDiffusionPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionMegaSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    tokenizer: CLIPTokenizer
+    image_feature_extractor: CLIPFeatureExtractor
+    text_encoder: CLIPTextModel
+    image_encoder: CLIPVisionModel
+    image_unet: UNet2DConditionModel
+    text_unet: UNet2DConditionModel
+    vae: AutoencoderKL
+    scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
+
+    def __init__(
+        self,
+        tokenizer: CLIPTokenizer,
+        image_feature_extractor: CLIPFeatureExtractor,
+        text_encoder: CLIPTextModel,
+        image_encoder: CLIPVisionModel,
+        image_unet: UNet2DConditionModel,
+        text_unet: UNet2DConditionModel,
+        vae: AutoencoderKL,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer,
+            image_feature_extractor=image_feature_extractor,
+            text_encoder=text_encoder,
+            image_encoder=image_encoder,
+            image_unet=image_unet,
+            text_unet=text_unet,
+            vae=vae,
+            scheduler=scheduler,
+        )
+
+    @property
+    def components(self) -> Dict[str, Any]:
+        return {k: getattr(self, k) for k in self.config.keys() if not k.startswith("_")}
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.image_unet.config.attention_head_dim // 2
+        self.image_unet.set_attention_slice(slice_size)
+        self.text_unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    @torch.no_grad()
+    def image_variation(
+        self,
+        image: Union[torch.FloatTensor, PIL.Image.Image],
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+    ):
+        return VersatileDiffusionImageVariationPipeline(**self.components)(
+            image=image,
+            height=height,
+            width=width,
+            num_inference_steps=num_inference_steps,
+            guidance_scale=guidance_scale,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            eta=eta,
+            generator=generator,
+            latents=latents,
+            output_type=output_type,
+            return_dict=return_dict,
+            callback=callback,
+            callback_steps=callback_steps,
+        )
+
+    @torch.no_grad()
+    def text_to_image(
+        self,
+        prompt: Union[str, List[str]],
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+    ):
+        return VersatileDiffusionTextToImagePipeline(**self.components)(
+            prompt=prompt,
+            height=height,
+            width=width,
+            num_inference_steps=num_inference_steps,
+            guidance_scale=guidance_scale,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            eta=eta,
+            generator=generator,
+            latents=latents,
+            output_type=output_type,
+            return_dict=return_dict,
+            callback=callback,
+            callback_steps=callback_steps,
+        )

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py

@@ -0,0 +1,602 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+
+import PIL
+from transformers import (
+    CLIPFeatureExtractor,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    CLIPVisionModelWithProjection,
+)
+
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention import DualTransformer2DModel, Transformer2DModel
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ...utils import is_accelerate_available, logging
+from .modeling_text_unet import UNetFlatConditionModel
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class VersatileDiffusionDualGuidedPipeline(DiffusionPipeline):
+    r"""
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Parameters:
+        vqvae ([`VQModel`]):
+            Vector-quantized (VQ) Model to encode and decode images to and from latent representations.
+        bert ([`LDMBertModel`]):
+            Text-encoder model based on [BERT](https://huggingface.co/docs/transformers/model_doc/bert) architecture.
+        tokenizer (`transformers.BertTokenizer`):
+            Tokenizer of class
+            [BertTokenizer](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+    """
+    tokenizer: CLIPTokenizer
+    image_feature_extractor: CLIPFeatureExtractor
+    text_encoder: CLIPTextModelWithProjection
+    image_encoder: CLIPVisionModelWithProjection
+    image_unet: UNet2DConditionModel
+    text_unet: UNetFlatConditionModel
+    vae: AutoencoderKL
+    scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
+
+    def __init__(
+        self,
+        tokenizer: CLIPTokenizer,
+        image_feature_extractor: CLIPFeatureExtractor,
+        text_encoder: CLIPTextModelWithProjection,
+        image_encoder: CLIPVisionModelWithProjection,
+        image_unet: UNet2DConditionModel,
+        text_unet: UNetFlatConditionModel,
+        vae: AutoencoderKL,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+    ):
+        super().__init__()
+        self.register_modules(
+            tokenizer=tokenizer,
+            image_feature_extractor=image_feature_extractor,
+            text_encoder=text_encoder,
+            image_encoder=image_encoder,
+            image_unet=image_unet,
+            text_unet=text_unet,
+            vae=vae,
+            scheduler=scheduler,
+        )
+
+    def convert_to_dual_attention(self, mix_ratio=0.5, condition_types=("image", "text")):
+        for name, module in self.image_unet.named_modules():
+            if isinstance(module, Transformer2DModel):
+                parent_name, index = name.rsplit(".", 1)
+                index = int(index)
+
+                image_transformer = self.image_unet.get_submodule(parent_name)[index]
+                text_transformer = self.text_unet.get_submodule(parent_name)[index]
+
+                config = image_transformer.config
+                dual_transformer = DualTransformer2DModel(
+                    num_attention_heads=config.num_attention_heads,
+                    attention_head_dim=config.attention_head_dim,
+                    in_channels=config.in_channels,
+                    num_layers=config.num_layers,
+                    dropout=config.dropout,
+                    norm_num_groups=config.norm_num_groups,
+                    cross_attention_dim=config.cross_attention_dim,
+                    attention_bias=config.attention_bias,
+                    sample_size=config.sample_size,
+                    num_vector_embeds=config.num_vector_embeds,
+                    activation_fn=config.activation_fn,
+                    num_embeds_ada_norm=config.num_embeds_ada_norm,
+                )
+                for i, type in enumerate(condition_types):
+                    if type == "image":
+                        dual_transformer.transformers[i] = image_transformer
+                    else:
+                        dual_transformer.transformers[i] = text_transformer
+
+                dual_transformer.mix_ratio = mix_ratio
+                self.image_unet.get_submodule(parent_name)[index] = dual_transformer
+
+    def remove_dual_attention(self):
+        for name, module in self.image_unet.named_modules():
+            if isinstance(module, DualTransformer2DModel):
+                parent_name, index = name.rsplit(".", 1)
+                index = int(index)
+                self.image_unet.get_submodule(parent_name)[index] = module.transformers[0]
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_xformers_memory_efficient_attention with unet->image_unet
+    def enable_xformers_memory_efficient_attention(self):
+        r"""
+        Enable memory efficient attention as implemented in xformers.
+
+        When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference
+        time. Speed up at training time is not guaranteed.
+
+        Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention
+        is used.
+        """
+        self.image_unet.set_use_memory_efficient_attention_xformers(True)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_xformers_memory_efficient_attention with unet->image_unet
+    def disable_xformers_memory_efficient_attention(self):
+        r"""
+        Disable memory efficient attention as implemented in xformers.
+        """
+        self.image_unet.set_use_memory_efficient_attention_xformers(False)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing with unet->image_unet
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.image_unet.config.attention_head_dim // 2
+        self.image_unet.set_attention_slice(slice_size)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_attention_slicing
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.image_unet, self.text_unet, self.text_encoder, self.vae]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device with unet->image_unet
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.image_unet, "_hf_hook"):
+            return self.device
+        for module in self.image_unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def _encode_text_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `list(int)`):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+        """
+
+        def normalize_embeddings(encoder_output):
+            embeds = self.text_encoder.text_projection(encoder_output.last_hidden_state)
+            embeds_pooled = encoder_output.text_embeds
+            embeds = embeds / torch.norm(embeds_pooled.unsqueeze(1), dim=-1, keepdim=True)
+            return embeds
+
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids
+
+        if not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+
+        if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+            attention_mask = text_inputs.attention_mask.to(device)
+        else:
+            attention_mask = None
+
+        text_embeddings = self.text_encoder(
+            text_input_ids.to(device),
+            attention_mask=attention_mask,
+        )
+        text_embeddings = normalize_embeddings(text_embeddings)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens = [""] * batch_size
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            uncond_embeddings = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            uncond_embeddings = normalize_embeddings(uncond_embeddings)
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        return text_embeddings
+
+    def _encode_image_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `list(int)`):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+        """
+
+        def normalize_embeddings(encoder_output):
+            embeds = self.image_encoder.vision_model.post_layernorm(encoder_output.last_hidden_state)
+            embeds = self.image_encoder.visual_projection(embeds)
+            embeds_pooled = embeds[:, 0:1]
+            embeds = embeds / torch.norm(embeds_pooled, dim=-1, keepdim=True)
+            return embeds
+
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        # get prompt text embeddings
+        image_input = self.image_feature_extractor(images=prompt, return_tensors="pt")
+        image_embeddings = self.image_encoder(image_input.pixel_values.to(device))
+        image_embeddings = normalize_embeddings(image_embeddings)
+
+        # duplicate image embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = image_embeddings.shape
+        image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1)
+        image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_images = [np.zeros((512, 512, 3))] * batch_size
+            uncond_images = self.image_feature_extractor(images=uncond_images, return_tensors="pt")
+            uncond_embeddings = self.image_encoder(uncond_images.pixel_values.to(device))
+            uncond_embeddings = normalize_embeddings(uncond_embeddings)
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and conditional embeddings into a single batch
+            # to avoid doing two forward passes
+            image_embeddings = torch.cat([uncond_embeddings, image_embeddings])
+
+        return image_embeddings
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+    def decode_latents(self, latents):
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(self, first_prompt, second_prompt, height, width, callback_steps):
+        if (
+            not isinstance(first_prompt, str)
+            and not isinstance(first_prompt, PIL.Image.Image)
+            and not isinstance(first_prompt, list)
+        ):
+            raise ValueError(
+                f"`first_prompt` has to be of type `str` `PIL.Image` or `list` but is {type(first_prompt)}"
+            )
+        if (
+            not isinstance(second_prompt, str)
+            and not isinstance(second_prompt, PIL.Image.Image)
+            and not isinstance(second_prompt, list)
+        ):
+            raise ValueError(
+                f"`second_prompt` has to be of type `str` `PIL.Image` or `list` but is {type(second_prompt)}"
+            )
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        if latents is None:
+            if device.type == "mps":
+                # randn does not work reproducibly on mps
+                latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device)
+            else:
+                latents = torch.randn(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            if latents.shape != shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def set_mix_ratio(self, mix_ratio):
+        for name, module in self.image_unet.named_modules():
+            if isinstance(module, DualTransformer2DModel):
+                module.mix_ratio = mix_ratio
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        first_prompt: Union[str, List[str], PIL.Image.Image, List[PIL.Image.Image]],
+        second_prompt: Union[str, List[str], PIL.Image.Image, List[PIL.Image.Image]],
+        prompt_mix_ratio: float = 0.5,
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(first_prompt, second_prompt, height, width, callback_steps)
+
+        # 2. Define call parameters
+        first_prompt = [first_prompt] if not isinstance(first_prompt, list) else first_prompt
+        second_prompt = [second_prompt] if not isinstance(second_prompt, list) else second_prompt
+        batch_size = len(first_prompt)
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompts
+        dual_prompt_embeddings = []
+        prompt_types = []
+        for prompt in [first_prompt, second_prompt]:
+            if isinstance(prompt[0], str):
+                embeddings = self._encode_text_prompt(
+                    prompt, device, num_images_per_prompt, do_classifier_free_guidance
+                )
+                prompt_types.append("text")
+            else:
+                embeddings = self._encode_image_prompt(
+                    prompt, device, num_images_per_prompt, do_classifier_free_guidance
+                )
+                prompt_types.append("image")
+            dual_prompt_embeddings.append(embeddings)
+        dual_prompt_embeddings = torch.cat(dual_prompt_embeddings, dim=1)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.image_unet.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            dual_prompt_embeddings.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Combine the attention blocks of the image and text UNets
+        self.convert_to_dual_attention(prompt_mix_ratio, prompt_types)
+        self.set_mix_ratio(prompt_mix_ratio)
+
+        # 8. Denoising loop
+        for i, t in enumerate(self.progress_bar(timesteps)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=dual_prompt_embeddings).sample
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        # 9. Return the image unet to its original state
+        self.remove_dual_attention()
+
+        # 10. Post-processing
+        image = self.decode_latents(latents)
+
+        # 11. Convert to PIL
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image,)
+
+        return ImagePipelineOutput(images=image)

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_image_to_text.py

@@ -0,0 +1,468 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from dataclasses import dataclass
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+
+import PIL
+from transformers import CLIPFeatureExtractor, CLIPVisionModelWithProjection, GPT2Tokenizer
+
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention import Transformer2DModel
+from ...pipeline_utils import BaseOutput, DiffusionPipeline
+from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ...utils import is_accelerate_available, logging
+from .modeling_gpt2_optimus import GPT2OptimusForLatentConnector
+from .modeling_text_unet import UNetFlatConditionModel
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class TextPipelineOutput(BaseOutput):
+    """
+    Output class for text generation pipelines.
+
+    Args:
+        text (`List[str]` or `np.ndarray`)
+            List of generated text of length `batch_size` or a numpy array of tokens of shape `(batch_size,
+            num_tokens)`.
+    """
+
+    text: Union[List[str], np.ndarray]
+
+
+class VersatileDiffusionImageToTextPipeline(DiffusionPipeline):
+    r"""
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Parameters:
+        vqvae ([`VQModel`]):
+            Vector-quantized (VQ) Model to encode and decode images to and from latent representations.
+        bert ([`LDMBertModel`]):
+            Text-encoder model based on [BERT](https://huggingface.co/docs/transformers/model_doc/bert) architecture.
+        tokenizer (`transformers.BertTokenizer`):
+            Tokenizer of class
+            [BertTokenizer](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+    """
+    image_feature_extractor: CLIPFeatureExtractor
+    image_encoder: CLIPVisionModelWithProjection
+    image_unet: UNet2DConditionModel
+    text_unet: UNetFlatConditionModel
+    vae: AutoencoderKL
+    scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
+
+    def __init__(
+        self,
+        image_feature_extractor: CLIPFeatureExtractor,
+        image_encoder: CLIPVisionModelWithProjection,
+        image_unet: UNet2DConditionModel,
+        text_unet: UNetFlatConditionModel,
+        vae: AutoencoderKL,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+    ):
+        super().__init__()
+        self.register_modules(
+            image_feature_extractor=image_feature_extractor,
+            image_encoder=image_encoder,
+            image_unet=image_unet,
+            text_unet=text_unet,
+            vae=vae,
+            scheduler=scheduler,
+        )
+
+        self.text_vae_decoder = GPT2OptimusForLatentConnector.from_pretrained("fusing/gpt2_optimus")
+        self.text_vae_tokenizer = GPT2Tokenizer.from_pretrained("fusing/gpt2_optimus")
+
+    def swap_unet_attention_blocks(self):
+        for name, module in self.image_unet.named_modules():
+            if isinstance(module, Transformer2DModel):
+                parent_name, index = name.rsplit(".", 1)
+                index = int(index)
+                self.image_unet.get_submodule(parent_name)[index], self.text_unet.get_submodule(parent_name)[index] = (
+                    self.text_unet.get_submodule(parent_name)[index],
+                    self.image_unet.get_submodule(parent_name)[index],
+                )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_xformers_memory_efficient_attention with unet->image_unet
+    def enable_xformers_memory_efficient_attention(self):
+        r"""
+        Enable memory efficient attention as implemented in xformers.
+
+        When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference
+        time. Speed up at training time is not guaranteed.
+
+        Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention
+        is used.
+        """
+        self.image_unet.set_use_memory_efficient_attention_xformers(True)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_xformers_memory_efficient_attention with unet->image_unet
+    def disable_xformers_memory_efficient_attention(self):
+        r"""
+        Disable memory efficient attention as implemented in xformers.
+        """
+        self.image_unet.set_use_memory_efficient_attention_xformers(False)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing with unet->image_unet
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.image_unet.config.attention_head_dim // 2
+        self.image_unet.set_attention_slice(slice_size)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_attention_slicing
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.image_unet, self.text_unet, self.text_encoder, self.vae]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device with unet->image_unet
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.image_unet, "_hf_hook"):
+            return self.device
+        for module in self.image_unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `list(int)`):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+        """
+
+        def normalize_embeddings(encoder_output):
+            embeds = self.image_encoder.vision_model.post_layernorm(encoder_output.last_hidden_state)
+            embeds = self.image_encoder.visual_projection(embeds)
+            embeds_pooled = embeds[:, 0:1]
+            embeds = embeds / torch.norm(embeds_pooled, dim=-1, keepdim=True)
+            return embeds
+
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        # get prompt text embeddings
+        # prompt = [(np.asarray(prompt) / 255)]
+        image_input = self.image_feature_extractor(images=prompt, return_tensors="pt")
+        image_embeddings = self.image_encoder(image_input.pixel_values.to(self.device))
+        image_embeddings = normalize_embeddings(image_embeddings)
+
+        # duplicate image embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = image_embeddings.shape
+        image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1)
+        image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_images: List[str]
+            if negative_prompt is None:
+                uncond_images = [np.zeros((512, 512, 3)) + 0.5] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, PIL.Image.Image):
+                uncond_images = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_images = negative_prompt
+
+            uncond_images = self.image_feature_extractor(images=uncond_images, return_tensors="pt")
+            uncond_embeddings = self.image_encoder(uncond_images.pixel_values.to(self.device))
+            uncond_embeddings = normalize_embeddings(uncond_embeddings)
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and conditional embeddings into a single batch
+            # to avoid doing two forward passes
+            image_embeddings = torch.cat([uncond_embeddings, image_embeddings])
+
+        return image_embeddings
+
+    def decode_latents(self, latents):
+        latents = latents.reshape(latents.shape[:-2])
+        self.text_vae_decoder = self.text_vae_decoder.to(self._execution_device)
+        bos_token = self.text_vae_tokenizer.bos_token_id
+        output = self.text_vae_decoder.generate(bos_token_id=bos_token, past=latents)
+        return output
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(self, image, callback_steps):
+        if not isinstance(image, PIL.Image.Image) and not isinstance(image, torch.Tensor):
+            raise ValueError(f"`image` has to be of type `PIL.Image.Image` or `torch.Tensor` but is {type(image)}")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+    def prepare_latents(self, batch_size, num_channels_latents, dtype, device, generator, latents=None):
+        shape = (batch_size, num_channels_latents, 1, 1)
+        if latents is None:
+            if device.type == "mps":
+                # randn does not work reproducibly on mps
+                latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device)
+            else:
+                latents = torch.randn(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            if latents.shape != shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.Tensor],
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "str",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            image (`PIL.Image.Image`, `List[PIL.Image.Image]` or `torch.Tensor`):
+                The image prompt or prompts to guide the image generation.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(image, callback_steps)
+
+        # 2. Define call parameters
+        batch_size = 1 if isinstance(image, PIL.Image.Image) else len(image)
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        image_embeddings = self._encode_prompt(
+            image, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
+        )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.text_unet.in_channels[0]
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            image_embeddings.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Swap the attention blocks between the image and text UNets
+        self.swap_unet_attention_blocks()
+
+        # 8. Denoising loop
+        for i, t in enumerate(self.progress_bar(timesteps)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            print("latent_model_input", latent_model_input.abs().sum())
+            print("timestep", t)
+
+            # predict the noise residual
+            noise_pred = self.text_unet(latent_model_input, t, encoder_hidden_states=image_embeddings).sample
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            print("e_t", noise_pred.abs().sum())
+            print("e_t[3,3]", noise_pred[0, :5, 0, 0])
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+            print("latents", latents.abs().sum())
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        # 9. Swap the attention blocks backs in case the UNets are reused in another pipeline
+        self.swap_unet_attention_blocks()
+
+        # 10. Post-processing
+        text = self.decode_latents(latents)
+
+        # 11. Convert to strings
+        if output_type == "str":
+            text = self.text_vae_tokenizer.batch_decode(text)
+
+        if not return_dict:
+            return (text,)
+
+        return TextPipelineOutput(text=text)

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py

@@ -0,0 +1,435 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+
+import PIL
+from transformers import CLIPFeatureExtractor, CLIPVisionModelWithProjection
+
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ...utils import is_accelerate_available, logging
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class VersatileDiffusionImageVariationPipeline(DiffusionPipeline):
+    r"""
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Parameters:
+        vqvae ([`VQModel`]):
+            Vector-quantized (VQ) Model to encode and decode images to and from latent representations.
+        bert ([`LDMBertModel`]):
+            Text-encoder model based on [BERT](https://huggingface.co/docs/transformers/model_doc/bert) architecture.
+        tokenizer (`transformers.BertTokenizer`):
+            Tokenizer of class
+            [BertTokenizer](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+    """
+    image_feature_extractor: CLIPFeatureExtractor
+    image_encoder: CLIPVisionModelWithProjection
+    image_unet: UNet2DConditionModel
+    vae: AutoencoderKL
+    scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
+
+    def __init__(
+        self,
+        image_feature_extractor: CLIPFeatureExtractor,
+        image_encoder: CLIPVisionModelWithProjection,
+        image_unet: UNet2DConditionModel,
+        vae: AutoencoderKL,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+    ):
+        super().__init__()
+        self.register_modules(
+            image_feature_extractor=image_feature_extractor,
+            image_encoder=image_encoder,
+            image_unet=image_unet,
+            vae=vae,
+            scheduler=scheduler,
+        )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_xformers_memory_efficient_attention with unet->image_unet
+    def enable_xformers_memory_efficient_attention(self):
+        r"""
+        Enable memory efficient attention as implemented in xformers.
+
+        When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference
+        time. Speed up at training time is not guaranteed.
+
+        Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention
+        is used.
+        """
+        self.image_unet.set_use_memory_efficient_attention_xformers(True)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_xformers_memory_efficient_attention with unet->image_unet
+    def disable_xformers_memory_efficient_attention(self):
+        r"""
+        Disable memory efficient attention as implemented in xformers.
+        """
+        self.image_unet.set_use_memory_efficient_attention_xformers(False)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing with unet->image_unet
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.image_unet.config.attention_head_dim // 2
+        self.image_unet.set_attention_slice(slice_size)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_attention_slicing
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.image_unet, self.text_unet, self.text_encoder, self.vae]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device with unet->image_unet
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.image_unet, "_hf_hook"):
+            return self.device
+        for module in self.image_unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `list(int)`):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+        """
+
+        def normalize_embeddings(encoder_output):
+            embeds = self.image_encoder.vision_model.post_layernorm(encoder_output.last_hidden_state)
+            embeds = self.image_encoder.visual_projection(embeds)
+            embeds_pooled = embeds[:, 0:1]
+            embeds = embeds / torch.norm(embeds_pooled, dim=-1, keepdim=True)
+            return embeds
+
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        # get prompt text embeddings
+        image_input = self.image_feature_extractor(images=prompt, return_tensors="pt")
+        image_embeddings = self.image_encoder(image_input.pixel_values.to(device))
+        image_embeddings = normalize_embeddings(image_embeddings)
+
+        # duplicate image embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = image_embeddings.shape
+        image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1)
+        image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_images: List[str]
+            if negative_prompt is None:
+                uncond_images = [np.zeros((512, 512, 3))] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, PIL.Image.Image):
+                uncond_images = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_images = negative_prompt
+
+            uncond_images = self.image_feature_extractor(images=uncond_images, return_tensors="pt")
+            uncond_embeddings = self.image_encoder(uncond_images.pixel_values.to(device))
+            uncond_embeddings = normalize_embeddings(uncond_embeddings)
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and conditional embeddings into a single batch
+            # to avoid doing two forward passes
+            image_embeddings = torch.cat([uncond_embeddings, image_embeddings])
+
+        return image_embeddings
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+    def decode_latents(self, latents):
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(self, image, height, width, callback_steps):
+        if not isinstance(image, PIL.Image.Image) and not isinstance(image, torch.Tensor):
+            raise ValueError(f"`image` has to be of type `PIL.Image.Image` or `torch.Tensor` but is {type(image)}")
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        if latents is None:
+            if device.type == "mps":
+                # randn does not work reproducibly on mps
+                latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device)
+            else:
+                latents = torch.randn(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            if latents.shape != shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.Tensor],
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            image (`PIL.Image.Image`, `List[PIL.Image.Image]` or `torch.Tensor`):
+                The image prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(image, height, width, callback_steps)
+
+        # 2. Define call parameters
+        batch_size = 1 if isinstance(image, PIL.Image.Image) else len(image)
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        image_embeddings = self._encode_prompt(
+            image, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
+        )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.image_unet.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            image_embeddings.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Denoising loop
+        for i, t in enumerate(self.progress_bar(timesteps)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=image_embeddings).sample
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        # 8. Post-processing
+        image = self.decode_latents(latents)
+
+        # 9. Convert to PIL
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image,)
+
+        return ImagePipelineOutput(images=image)

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py

@@ -0,0 +1,494 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Union
+
+import torch
+import torch.utils.checkpoint
+
+from transformers import CLIPFeatureExtractor, CLIPTextModelWithProjection, CLIPTokenizer
+
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention import Transformer2DModel
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ...utils import is_accelerate_available, logging
+from .modeling_text_unet import UNetFlatConditionModel
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class VersatileDiffusionTextToImagePipeline(DiffusionPipeline):
+    r"""
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Parameters:
+        vqvae ([`VQModel`]):
+            Vector-quantized (VQ) Model to encode and decode images to and from latent representations.
+        bert ([`LDMBertModel`]):
+            Text-encoder model based on [BERT](https://huggingface.co/docs/transformers/model_doc/bert) architecture.
+        tokenizer (`transformers.BertTokenizer`):
+            Tokenizer of class
+            [BertTokenizer](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+    """
+    tokenizer: CLIPTokenizer
+    image_feature_extractor: CLIPFeatureExtractor
+    text_encoder: CLIPTextModelWithProjection
+    image_unet: UNet2DConditionModel
+    text_unet: UNetFlatConditionModel
+    vae: AutoencoderKL
+    scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
+
+    def __init__(
+        self,
+        tokenizer: CLIPTokenizer,
+        text_encoder: CLIPTextModelWithProjection,
+        image_unet: UNet2DConditionModel,
+        text_unet: UNetFlatConditionModel,
+        vae: AutoencoderKL,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+    ):
+        super().__init__()
+        self.register_modules(
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+            image_unet=image_unet,
+            text_unet=text_unet,
+            vae=vae,
+            scheduler=scheduler,
+        )
+
+    def swap_unet_attention_blocks(self):
+        for name, module in self.image_unet.named_modules():
+            if isinstance(module, Transformer2DModel):
+                parent_name, index = name.rsplit(".", 1)
+                index = int(index)
+                self.image_unet.get_submodule(parent_name)[index], self.text_unet.get_submodule(parent_name)[index] = (
+                    self.text_unet.get_submodule(parent_name)[index],
+                    self.image_unet.get_submodule(parent_name)[index],
+                )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_xformers_memory_efficient_attention with unet->image_unet
+    def enable_xformers_memory_efficient_attention(self):
+        r"""
+        Enable memory efficient attention as implemented in xformers.
+
+        When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference
+        time. Speed up at training time is not guaranteed.
+
+        Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention
+        is used.
+        """
+        self.image_unet.set_use_memory_efficient_attention_xformers(True)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_xformers_memory_efficient_attention with unet->image_unet
+    def disable_xformers_memory_efficient_attention(self):
+        r"""
+        Disable memory efficient attention as implemented in xformers.
+        """
+        self.image_unet.set_use_memory_efficient_attention_xformers(False)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing with unet->image_unet
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.image_unet.config.attention_head_dim // 2
+        self.image_unet.set_attention_slice(slice_size)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_attention_slicing
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.image_unet, self.text_unet, self.text_encoder, self.vae]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device with unet->image_unet
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.image_unet, "_hf_hook"):
+            return self.device
+        for module in self.image_unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `list(int)`):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+        """
+
+        def normalize_embeddings(encoder_output):
+            embeds = self.text_encoder.text_projection(encoder_output.last_hidden_state)
+            embeds_pooled = encoder_output.text_embeds
+            embeds = embeds / torch.norm(embeds_pooled.unsqueeze(1), dim=-1, keepdim=True)
+            return embeds
+
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids
+
+        if not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+
+        if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+            attention_mask = text_inputs.attention_mask.to(device)
+        else:
+            attention_mask = None
+
+        text_embeddings = self.text_encoder(
+            text_input_ids.to(device),
+            attention_mask=attention_mask,
+        )
+        text_embeddings = normalize_embeddings(text_embeddings)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            uncond_embeddings = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            uncond_embeddings = normalize_embeddings(uncond_embeddings)
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        return text_embeddings
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+    def decode_latents(self, latents):
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs
+    def check_inputs(self, prompt, height, width, callback_steps):
+        if not isinstance(prompt, str) and not isinstance(prompt, list):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        if latents is None:
+            if device.type == "mps":
+                # randn does not work reproducibly on mps
+                latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device)
+            else:
+                latents = torch.randn(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            if latents.shape != shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(prompt, height, width, callback_steps)
+
+        # 2. Define call parameters
+        batch_size = 1 if isinstance(prompt, str) else len(prompt)
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        text_embeddings = self._encode_prompt(
+            prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
+        )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.image_unet.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            text_embeddings.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Swap the attention blocks between the image and text UNets
+        self.swap_unet_attention_blocks()
+
+        # 8. Denoising loop
+        for i, t in enumerate(self.progress_bar(timesteps)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        # 9. Swap the attention blocks backs in case the UNets are reused in another pipeline
+        self.swap_unet_attention_blocks()
+
+        # 10. Post-processing
+        image = self.decode_latents(latents)
+
+        # 11. Convert to PIL
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image,)
+
+        return ImagePipelineOutput(images=image)

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -124,6 +124,51 @@ class StableDiffusionPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class VersatileDiffusionImageVariationPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class VersatileDiffusionPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class VersatileDiffusionTextToImagePipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class VQDiffusionPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/versatile_diffusion/test_versatile_diffusion_dual_guided.py

@@ -0,0 +1,104 @@
+# coding=utf-8
+# Copyright 2022 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import gc
+import tempfile
+import unittest
+
+import numpy as np
+import torch
+
+from diffusers import VersatileDiffusionDualGuidedPipeline
+from diffusers.utils.testing_utils import load_image, require_torch_gpu, slow, torch_device
+
+from ...test_pipelines_common import PipelineTesterMixin
+
+
+torch.backends.cuda.matmul.allow_tf32 = False
+
+
+class VersatileDiffusionDualGuidedPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pass
+
+
+@slow
+@require_torch_gpu
+class VersatileDiffusionDualGuidedPipelineIntegrationTests(unittest.TestCase):
+    def tearDown(self):
+        # clean up the VRAM after each test
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_from_pretrained_save_pretrained(self):
+        pipe = VersatileDiffusionDualGuidedPipeline.from_pretrained("diffusers/vd-official-test")
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        image = pipe(
+            first_prompt="first prompt",
+            second_prompt="second prompt",
+            prompt_mix_ratio=0.75,
+            generator=generator,
+            guidance_scale=7.5,
+            num_inference_steps=2,
+            output_type="numpy",
+        ).images
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            pipe.save_pretrained(tmpdirname)
+            pipe = VersatileDiffusionDualGuidedPipeline.from_pretrained(tmpdirname)
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator = generator.manual_seed(0)
+        new_image = pipe(
+            first_prompt="first prompt",
+            second_prompt="second prompt",
+            prompt_mix_ratio=0.75,
+            generator=generator,
+            guidance_scale=7.5,
+            num_inference_steps=2,
+            output_type="numpy",
+        ).images
+
+        assert np.abs(image - new_image).sum() < 1e-5, "Models don't have the same forward pass"
+
+    def test_inference_image_variations(self):
+        pipe = VersatileDiffusionDualGuidedPipeline.from_pretrained("diffusers/vd-official-test")
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        first_prompt = "cyberpunk 2077"
+        second_prompt = load_image(
+            "https://raw.githubusercontent.com/SHI-Labs/Versatile-Diffusion/master/assets/benz.jpg"
+        )
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        image = pipe(
+            first_prompt=first_prompt,
+            second_prompt=second_prompt,
+            prompt_mix_ratio=0.75,
+            generator=generator,
+            guidance_scale=7.5,
+            num_inference_steps=50,
+            output_type="numpy",
+        ).images
+
+        image_slice = image[0, 253:256, 253:256, -1]
+
+        assert image.shape == (1, 512, 512, 3)
+        expected_slice = np.array([0.1811, 0.0430, 0.0433, 0.1082, 0.0144, 0.0306, 0.0683, 0.0248, 0.0876])
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

tests/pipelines/versatile_diffusion/test_versatile_diffusion_image_to_text.py

@@ -0,0 +1,57 @@
+# coding=utf-8
+# Copyright 2022 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import numpy as np
+import torch
+
+from diffusers import VersatileDiffusionImageToTextPipeline, DDIMScheduler
+from diffusers.utils.testing_utils import load_image, require_torch_gpu, slow, torch_device
+
+from ...test_pipelines_common import PipelineTesterMixin
+
+
+torch.backends.cuda.matmul.allow_tf32 = False
+
+
+class VersatileDiffusionImageToTextPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pass
+
+
+@slow
+@require_torch_gpu
+class VersatileDiffusionImageToTextPipelineIntegrationTests(unittest.TestCase):
+    def test_inference_image_to_text(self):
+        pipe = VersatileDiffusionImageToTextPipeline.from_pretrained("diffusers/vd-official-test")
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        image_prompt = load_image(
+            "https://raw.githubusercontent.com/SHI-Labs/Versatile-Diffusion/master/assets/boy_and_girl.jpg"
+        )
+        # generator = torch.Generator(device=torch_device).manual_seed(0)
+        np.random.seed(8)
+        torch.manual_seed(108)
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        text = pipe(
+            image=image_prompt,
+            # generator=generator,
+            guidance_scale=7.5,
+            num_inference_steps=50,
+            output_type="str",
+        ).text
+
+        assert text == "Corret me"

tests/pipelines/versatile_diffusion/test_versatile_diffusion_image_variation.py

@@ -0,0 +1,58 @@
+# coding=utf-8
+# Copyright 2022 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import numpy as np
+import torch
+
+from diffusers import VersatileDiffusionImageVariationPipeline
+from diffusers.utils.testing_utils import load_image, require_torch_gpu, slow, torch_device
+
+from ...test_pipelines_common import PipelineTesterMixin
+
+
+torch.backends.cuda.matmul.allow_tf32 = False
+
+
+class VersatileDiffusionImageVariationPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pass
+
+
+@slow
+@require_torch_gpu
+class VersatileDiffusionImageVariationPipelineIntegrationTests(unittest.TestCase):
+    def test_inference_image_variations(self):
+        pipe = VersatileDiffusionImageVariationPipeline.from_pretrained("diffusers/vd-official-test")
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        image_prompt = load_image(
+            "https://raw.githubusercontent.com/SHI-Labs/Versatile-Diffusion/master/assets/benz.jpg"
+        )
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        image = pipe(
+            image=image_prompt,
+            generator=generator,
+            guidance_scale=7.5,
+            num_inference_steps=50,
+            output_type="numpy",
+        ).images
+
+        image_slice = image[0, 253:256, 253:256, -1]
+
+        assert image.shape == (1, 512, 512, 3)
+        expected_slice = np.array([0.1811, 0.0430, 0.0433, 0.1082, 0.0144, 0.0306, 0.0683, 0.0248, 0.0876])
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

tests/pipelines/versatile_diffusion/test_versatile_diffusion_text_to_image.py

@@ -0,0 +1,52 @@
+# coding=utf-8
+# Copyright 2022 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import numpy as np
+import torch
+
+from diffusers import VersatileDiffusionTextToImagePipeline
+from diffusers.utils.testing_utils import require_torch_gpu, slow, torch_device
+
+from ...test_pipelines_common import PipelineTesterMixin
+
+
+torch.backends.cuda.matmul.allow_tf32 = False
+
+
+class VersatileDiffusionTextToImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pass
+
+
+@slow
+@require_torch_gpu
+class VersatileDiffusionTextToImagePipelineIntegrationTests(unittest.TestCase):
+    def test_inference_text2img(self):
+        pipe = VersatileDiffusionTextToImagePipeline.from_pretrained("diffusers/vd-official-test")
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        prompt = "A painting of a squirrel eating a burger "
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        image = pipe(
+            prompt=prompt, generator=generator, guidance_scale=7.5, num_inference_steps=50, output_type="numpy"
+        ).images
+
+        image_slice = image[0, 253:256, 253:256, -1]
+
+        assert image.shape == (1, 512, 512, 3)
+        expected_slice = np.array([0.0657, 0.0529, 0.0455, 0.0802, 0.0570, 0.0179, 0.0267, 0.0483, 0.0769])
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2