frame wise guidance

scripts/convert_open_sora_stdit_to_diffusers.py

@@ -0,0 +1,172 @@
+import argparse
+import os
+
+import torch
+from safetensors import safe_open
+
+from diffusers import Transformer3DModel
+
+
+ckpt_id = "PixArt-alpha/PixArt-alpha"
+# https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/scripts/inference.py#L125
+interpolation_scale = {256: 0.5, 512: 1}
+
+
+def main(args):
+    state_dict = {}
+    with safe_open(args.orig_ckpt_path, framework="pt", device="cpu") as f:
+        for k in f.keys():
+            state_dict[k] = f.get_tensor(k)
+    converted_state_dict = {}
+
+    # Patch embeddings.
+    converted_state_dict["pos_embed.proj.weight"] = state_dict.pop("x_embedder.proj.weight")
+    converted_state_dict["pos_embed.proj.bias"] = state_dict.pop("x_embedder.proj.bias")
+
+    # Caption projection.
+    converted_state_dict["caption_projection.linear_1.weight"] = state_dict.pop("y_embedder.y_proj.fc1.weight")
+    converted_state_dict["caption_projection.linear_1.bias"] = state_dict.pop("y_embedder.y_proj.fc1.bias")
+    converted_state_dict["caption_projection.linear_2.weight"] = state_dict.pop("y_embedder.y_proj.fc2.weight")
+    converted_state_dict["caption_projection.linear_2.bias"] = state_dict.pop("y_embedder.y_proj.fc2.bias")
+
+    # AdaLN-single LN
+    converted_state_dict["adaln_single.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
+        "t_embedder.mlp.0.weight"
+    )
+    converted_state_dict["adaln_single.emb.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
+    converted_state_dict["adaln_single.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
+        "t_embedder.mlp.2.weight"
+    )
+    converted_state_dict["adaln_single.emb.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
+
+    # Shared norm.
+    converted_state_dict["adaln_single.linear.weight"] = state_dict.pop("t_block.1.weight")
+    converted_state_dict["adaln_single.linear.bias"] = state_dict.pop("t_block.1.bias")
+
+    for depth in range(28):
+        # Transformer blocks.
+        converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
+            f"blocks.{depth}.scale_shift_table"
+        )
+
+        # Attention is all you need 🤘
+
+        # Self attention.
+        q, k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.attn.qkv.weight"), 3, dim=0)
+        q_bias, k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.attn.qkv.bias"), 3, dim=0)
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.bias"] = v_bias
+        # Projection.
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.attn.proj.bias"
+        )
+
+        # Temporal attention.
+        q, k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.attn_temp.qkv.weight"), 3, dim=0)
+        q_bias, k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.attn_temp.qkv.bias"), 3, dim=0)
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_v.bias"] = v_bias
+        # Projection.
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.attn_temp.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.attn_temp.proj.bias"
+        )
+
+        # Feed-forward.
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.0.proj.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc1.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.0.proj.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc1.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.2.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc2.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.2.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc2.bias"
+        )
+
+        # Cross-attention.
+        q = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.weight")
+        q_bias = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.bias")
+        k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.weight"), 2, dim=0)
+        k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.bias"), 2, dim=0)
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.bias"] = v_bias
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.cross_attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.cross_attn.proj.bias"
+        )
+
+    # Final block.
+    converted_state_dict["proj_out.weight"] = state_dict.pop("final_layer.linear.weight")
+    converted_state_dict["proj_out.bias"] = state_dict.pop("final_layer.linear.bias")
+    converted_state_dict["scale_shift_table"] = state_dict.pop("final_layer.scale_shift_table")
+    converted_state_dict["pos_embed_temporal"] = state_dict.pop("pos_embed_temporal")
+
+    # DiT XL/2
+    transformer = Transformer3DModel(
+        sample_size=(16, args.image_size // 8, args.image_size // 8),
+        patch_size=(1, 2, 2),
+        num_layers=28,
+        attention_head_dim=72,
+        num_attention_heads=16,
+        in_channels=4,
+        out_channels=8,
+        cross_attention_dim=1152,
+        num_embeds_ada_norm=1000,
+        norm_eps=1e-6,
+        caption_channels=4096,
+    )
+    transformer.load_state_dict(converted_state_dict, strict=True)
+
+    assert transformer.pos_embed.pos_embed is not None
+    state_dict.pop("pos_embed")
+    state_dict.pop("y_embedder.y_embedding")
+    assert len(state_dict) == 0, f"State dict is not empty, {state_dict.keys()}"
+
+    num_model_params = sum(p.numel() for p in transformer.parameters())
+    print(f"Total number of transformer parameters: {num_model_params}")
+
+    transformer.save_pretrained(os.path.join(args.dump_path, "transformer"))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--orig_ckpt_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--image_size",
+        default=256,
+        type=int,
+        choices=[256, 512],
+        required=False,
+        help="Image size of pretrained model, either 256 or 512.",
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
+
+    args = parser.parse_args()
+    main(args)

src/diffusers/__init__.py

@@ -90,6 +90,7 @@ else:
             "T2IAdapter",
             "T5FilmDecoder",
             "Transformer2DModel",
+            "Transformer3DModel",
             "UNet1DModel",
             "UNet2DConditionModel",
             "UNet2DModel",
@@ -256,6 +257,7 @@ else:
             "LEditsPPPipelineStableDiffusion",
             "LEditsPPPipelineStableDiffusionXL",
             "MusicLDMPipeline",
+            "OpenSoraPipeline",
             "PaintByExamplePipeline",
             "PIAPipeline",
             "PixArtAlphaPipeline",
@@ -483,6 +485,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             T2IAdapter,
             T5FilmDecoder,
             Transformer2DModel,
+            Transformer3DModel,
             UNet1DModel,
             UNet2DConditionModel,
             UNet2DModel,
@@ -628,6 +631,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LEditsPPPipelineStableDiffusion,
             LEditsPPPipelineStableDiffusionXL,
             MusicLDMPipeline,
+            OpenSoraPipeline,
             PaintByExamplePipeline,
             PIAPipeline,
             PixArtAlphaPipeline,

src/diffusers/models/__init__.py

@@ -38,6 +38,7 @@ if is_torch_available():
     _import_structure["transformers.prior_transformer"] = ["PriorTransformer"]
     _import_structure["transformers.t5_film_transformer"] = ["T5FilmDecoder"]
     _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
+    _import_structure["transformers.transformer_3d"] = ["Transformer3DModel"]
     _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
     _import_structure["unets.unet_1d"] = ["UNet1DModel"]
     _import_structure["unets.unet_2d"] = ["UNet2DModel"]
@@ -75,6 +76,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             PriorTransformer,
             T5FilmDecoder,
             Transformer2DModel,
+            Transformer3DModel,
             TransformerTemporalModel,
         )
         from .unets import (

src/diffusers/models/embeddings.py

@@ -187,6 +187,75 @@ class PatchEmbed(nn.Module):
         return (latent + pos_embed).to(latent.dtype)
 
 
+class PatchEmbed3D(nn.Module):
+    """Video to Patch Embedding"""
+
+    def __init__(
+        self,
+        height=224,
+        width=224,
+        patch_size=(1, 2, 2),
+        in_channels=3,
+        embed_dim=768,
+        layer_norm=False,
+        bias=True,
+        interpolation_scale=1,
+    ):
+        super().__init__()
+
+        num_patches = (height // patch_size[1]) * (width // patch_size[2])
+        self.layer_norm = layer_norm
+        self.emed_dim = embed_dim
+
+        self.proj = nn.Conv3d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
+        if layer_norm:
+            self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6)
+        else:
+            self.norm = None
+
+        self.patch_size = patch_size
+        # See:
+        # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L161
+        self.height, self.width = height // patch_size[1], width // patch_size[2]
+        self.base_size = height // patch_size[1]
+        self.interpolation_scale = interpolation_scale
+        pos_embed = get_2d_sincos_pos_embed(
+            embed_dim, int(num_patches**0.5), base_size=self.base_size, interpolation_scale=self.interpolation_scale
+        )
+        self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)
+
+    def forward(self, latent):
+        height, width = latent.shape[-2] // self.patch_size[1], latent.shape[-1] // self.patch_size[2]
+
+        latent = self.proj(latent)  # (B C T H W)
+
+        if self.layer_norm:
+            batch_size, _, num_frames, height, width = latent.size()
+            latent = latent.flatten(2).transpose(1, 2)
+            latent = self.norm(latent)
+            latent = latent.transpose(1, 2).view(batch_size, self.emed_dim, num_frames, height, width)
+
+        latent = latent.flatten(3).permute(0, 2, 3, 1)  # BCTHW -> BT(HW)C
+
+        # Interpolate positional embeddings if needed.
+        # (For PixArt-Alpha: https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L162C151-L162C160)
+        if self.height != height or self.width != width:
+            pos_embed = get_2d_sincos_pos_embed(
+                embed_dim=self.pos_embed.shape[-1],
+                grid_size=(height, width),
+                base_size=self.base_size,
+                interpolation_scale=self.interpolation_scale,
+            )
+            pos_embed = torch.from_numpy(pos_embed)
+            pos_embed = pos_embed.float().unsqueeze(0).to(latent.device)
+        else:
+            pos_embed = self.pos_embed
+
+        latent = (latent + pos_embed).to(latent.dtype)
+        latent = latent.flatten(1, 2)  # BT(H*W)C -> B(T*H*W)C
+        return latent
+
+
 class TimestepEmbedding(nn.Module):
     def __init__(
         self,

src/diffusers/models/transformers/__init__.py

@@ -6,4 +6,5 @@ if is_torch_available():
     from .prior_transformer import PriorTransformer
     from .t5_film_transformer import T5FilmDecoder
     from .transformer_2d import Transformer2DModel
+    from .transformer_3d import Transformer3DModel
     from .transformer_temporal import TransformerTemporalModel

src/diffusers/models/transformers/transformer_3d.py

@@ -0,0 +1,434 @@
+# Copyright 2024 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.
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import numpy as np
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput, logging
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import FeedForward, _chunked_feed_forward
+from ..attention_processor import Attention
+from ..embeddings import PatchEmbed3D, PixArtAlphaTextProjection, get_1d_sincos_pos_embed_from_grid
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormSingle
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class Transformer3DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer3DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+
+
+def get_1d_sincos_pos_embed(embed_dim, length, scale=1.0):
+    pos = np.arange(0, length)[..., None] / scale
+    return get_1d_sincos_pos_embed_from_grid(embed_dim, pos)
+
+
+@maybe_allow_in_graph
+class Transformer3DBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        cross_attention_dim: int,
+        num_embeds_ada_norm: Optional[int] = None,
+        norm_eps: float = 1e-6,
+        num_temporal_patches: int = 16,
+        num_spatial_patches: int = 256,
+    ):
+        super().__init__()
+        # We keep these boolean flags for backward-compatibility.
+
+        self.use_ada_layer_norm_single = True
+        self.num_embeds_ada_norm = num_embeds_ada_norm
+        self.num_temporal_patches = num_temporal_patches
+        self.num_spatial_patches = num_spatial_patches
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Spatial Self-Attn
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=norm_eps)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            bias=True,
+        )
+
+        # 2. Temporal Self-Attn
+        self.attn_temporal = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            bias=True,
+        )
+
+        # 2. Cross-Attn
+        # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+        # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+        # the second cross attention block.
+        self.norm2 = nn.LayerNorm(dim, norm_eps, elementwise_affine=False)
+        self.attn2 = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            bias=True,
+        )
+
+        # 3. Feed-forward
+        self.ff = FeedForward(dim, activation_fn="gelu-approximate")
+
+        # 4. Scale-shift for PixArt-Alpha.
+        self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor,
+        timestep: torch.LongTensor,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        temporal_pos_embed: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+            self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+        ).chunk(6, dim=1)
+        norm_hidden_states = self.norm1(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+        norm_hidden_states = norm_hidden_states.squeeze(1)
+
+        # 1. Spatial Self-Attention
+        # reshape (batch, num_temporal_patches*num_spatial_patches, dim) -> (batch * num_temporal_patches, num_spatial_patches, dim)
+        norm_hidden_states = norm_hidden_states.view(
+            batch_size, self.num_temporal_patches, self.num_spatial_patches, -1
+        )
+        norm_hidden_states = norm_hidden_states.view(
+            batch_size * self.num_temporal_patches, self.num_spatial_patches, -1
+        )
+
+        attn_output = self.attn1(norm_hidden_states)
+
+        # reshape (batch * num_temporal_patches, num_spatial_patches, dim) -> (batch, num_temporal_patches*num_spatial_patches, dim)
+        attn_output = attn_output.view(batch_size, self.num_temporal_patches, self.num_spatial_patches, -1)
+        attn_output = attn_output.view(batch_size, self.num_temporal_patches * self.num_spatial_patches, -1)
+
+        attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2. Temporal Self-Attention
+        # reshape (batch, num_temporal_patches*num_spatial_patches, dim) -> (batch * num_spatial_patches, num_temporal_patches, dim)
+        temporal_hidden_states = (
+            hidden_states.view(batch_size, self.num_temporal_patches, self.num_spatial_patches, -1)
+            .transpose(1, 2)
+            .contiguous()
+        )
+        temporal_hidden_states = temporal_hidden_states.view(
+            batch_size * self.num_spatial_patches, self.num_temporal_patches, -1
+        )
+
+        if temporal_pos_embed is not None:
+            temporal_hidden_states = temporal_hidden_states + temporal_pos_embed
+
+        attn_output = self.attn_temporal(temporal_hidden_states)
+
+        # reshape (batch * num_spatial_patches, num_temporal_patches, dim) -> (batch, num_temporal_patches*num_spatial_patches, dim)
+        attn_output = (
+            attn_output.view(batch_size, self.num_spatial_patches, self.num_temporal_patches, -1)
+            .transpose(1, 2)
+            .contiguous()
+        )
+        attn_output = attn_output.view(batch_size, self.num_temporal_patches * self.num_spatial_patches, -1)
+
+        attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 3. Cross-Attention
+        # For PixArt norm2 isn't applied here:
+        # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+        norm_hidden_states = hidden_states
+        attn_output = self.attn2(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=encoder_attention_mask,
+        )
+        hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class Transformer3DModel(ModelMixin, ConfigMixin):
+    """
+    A 3D Transformer model for image-like data.
+
+    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*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since 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 steps than `num_embeds_ada_norm`.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        sample_size: Tuple[int] = (2, 4, 4),
+        patch_size: Tuple[int] = (1, 2, 2),
+        in_channels: int = 4,
+        out_channels: int = 8,
+        num_layers: int = 1,
+        cross_attention_dim: int = 256,
+        num_embeds_ada_norm: int = 1000,
+        norm_eps: float = 1e-6,
+        caption_channels: int = 256,
+        interpolation_scale: float = None,
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # 2. Define input layers
+        self.height = sample_size[1]
+        self.width = sample_size[2]
+        self.num_patches = np.prod([sample_size[i] // patch_size[i] for i in range(3)])
+        self.num_temporal_patches = sample_size[0] // patch_size[0]
+        self.num_spatial_patches = self.num_patches // self.num_temporal_patches
+
+        self.patch_size = patch_size
+        interpolation_scale = (
+            interpolation_scale if interpolation_scale is not None else max(self.config.sample_size[1] // 64, 1)
+        )
+        self.pos_embed = PatchEmbed3D(
+            height=self.height,
+            width=self.width,
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=inner_dim,
+            interpolation_scale=interpolation_scale,
+        )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                Transformer3DBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    cross_attention_dim=cross_attention_dim,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    norm_eps=norm_eps,
+                    num_temporal_patches=self.num_temporal_patches,
+                    num_spatial_patches=self.num_spatial_patches,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+        self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+        self.proj_out = nn.Linear(inner_dim, np.prod(patch_size) * self.out_channels)
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = AdaLayerNormSingle(inner_dim)
+
+        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+
+        interpolation_scale = max(self.config.sample_size[0] // 16, 1)
+        temporal_pos_embed = get_1d_sincos_pos_embed(inner_dim, self.num_temporal_patches, scale=interpolation_scale)
+        temporal_pos_embed = torch.from_numpy(temporal_pos_embed).float().unsqueeze(0).requires_grad_(False)
+        self.register_buffer("pos_embed_temporal", temporal_pos_embed)
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        height, width = hidden_states.shape[-2] // self.patch_size[1], hidden_states.shape[-1] // self.patch_size[2]
+        # import ipdb; ipdb.set_trace()
+        hidden_states = self.pos_embed(hidden_states)
+
+        batch_size = hidden_states.shape[0]
+        added_cond_kwargs = {"resolution": None, "aspect_ratio": None}
+        timestep, embedded_timestep = self.adaln_single(
+            timestep, added_cond_kwargs=added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+        )
+
+        # 2. Blocks
+        encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        for i, block in enumerate(self.transformer_blocks):
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                timestep=timestep,
+                temporal_pos_embed=self.pos_embed_temporal if i == 0 else None,
+            )
+
+        # 3. Output
+        shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
+        hidden_states = self.norm_out(hidden_states)
+        # Modulation
+        hidden_states = hidden_states * (1 + scale) + shift
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = hidden_states.squeeze(1)
+
+        # unpatchify
+        hidden_states = hidden_states.reshape(
+            shape=(
+                -1,
+                self.num_temporal_patches,
+                height,
+                width,
+                self.patch_size[1],
+                self.patch_size[2],
+                self.out_channels,
+            )
+        )
+        hidden_states = torch.einsum("nthwpqc->ncthpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(
+                -1,
+                self.out_channels,
+                self.num_temporal_patches,
+                height * self.patch_size[1],
+                width * self.patch_size[2],
+            )
+        )
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer3DModelOutput(sample=output)

src/diffusers/pipelines/__init__.py

@@ -247,6 +247,7 @@ else:
         "WuerstchenDecoderPipeline",
         "WuerstchenPriorPipeline",
     ]
+    _import_structure["open_sora"] = ["OpenSoraPipeline"]
 try:
     if not is_onnx_available():
         raise OptionalDependencyNotAvailable()
@@ -438,6 +439,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LEditsPPPipelineStableDiffusionXL,
         )
         from .musicldm import MusicLDMPipeline
+        from .open_sora import OpenSoraPipeline
         from .paint_by_example import PaintByExamplePipeline
         from .pia import PIAPipeline
         from .pixart_alpha import PixArtAlphaPipeline

src/diffusers/pipelines/open_sora/__init__.py

@@ -0,0 +1,48 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_open_sora"] = ["OpenSoraPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_open_sora import OpenSoraPipeline
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/pipelines/open_sora/pipeline_open_sora.py

@@ -0,0 +1,912 @@
+# Copyright 2024 Open-Sora Authors and 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 html
+import inspect
+import re
+import urllib.parse as ul
+from dataclasses import dataclass
+from typing import Callable, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...models import AutoencoderKL, Transformer3DModel
+from ...schedulers import DPMSolverMultistepScheduler
+from ...utils import (
+    BACKENDS_MAPPING,
+    BaseOutput,
+    deprecate,
+    is_bs4_available,
+    is_ftfy_available,
+    logging,
+    replace_example_docstring,
+)
+from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+if is_bs4_available():
+    from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+    import ftfy
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import OpenSoraPipeline
+
+        >>> # You can replace the checkpoint id with "PixArt-alpha/PixArt-XL-2-512x512" too.
+        >>> pipe = OpenSoraPipeline.from_pretrained("PixArt-alpha/PixArt-XL-2-1024-MS", torch_dtype=torch.float16)
+        >>> # Enable memory optimizations.
+        >>> pipe.enable_model_cpu_offload()
+
+        >>> prompt = "A small cactus with a happy face in the Sahara desert."
+        >>> image = pipe(prompt).images[0]
+        ```
+"""
+
+
+@dataclass
+class OpenSoraPipelineOutput(BaseOutput):
+    r"""
+    Output class for Stable Video Diffusion pipeline.
+
+    Args:
+        frames (`[List[List[PIL.Image.Image]]`, `np.ndarray`, `torch.FloatTensor`]):
+            List of denoised PIL images of length `batch_size` or numpy array or torch tensor
+            of shape `(batch_size, num_frames, height, width, num_channels)`.
+    """
+
+    frames: Union[List[List[PIL.Image.Image]], np.ndarray, torch.FloatTensor]
+
+
+def _append_dims(x, target_dims):
+    """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
+    dims_to_append = target_dims - x.ndim
+    if dims_to_append < 0:
+        raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less")
+    return x[(...,) + (None,) * dims_to_append]
+
+
+# Copied from diffusers.pipelines.animatediff.pipeline_animatediff.tensor2vid
+def tensor2vid(video: torch.Tensor, processor: VaeImageProcessor, output_type: str = "np"):
+    batch_size, channels, num_frames, height, width = video.shape
+    outputs = []
+    for batch_idx in range(batch_size):
+        batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+        batch_output = processor.postprocess(batch_vid, output_type)
+
+        outputs.append(batch_output)
+
+    if output_type == "np":
+        outputs = np.stack(outputs)
+
+    elif output_type == "pt":
+        outputs = torch.stack(outputs)
+
+    elif not output_type == "pil":
+        raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']")
+
+    return outputs
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used,
+            `timesteps` must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+                Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+                timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+                must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class OpenSoraPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Open-Sora.
+
+    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 ([`T5EncoderModel`]):
+            Frozen text-encoder. Open-Sora uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the
+            [t5-v1_1-xxl](https://huggingface.co/Open-Sora/Open-Sora/tree/main/t5-v1_1-xxl) variant.
+        tokenizer (`T5Tokenizer`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`Transformer2DModel`]):
+            A text conditioned `Transformer2DModel` to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+    """
+
+    bad_punct_regex = re.compile(
+        r"["
+        + "#®•©™&@·º½¾¿¡§~"
+        + r"\)"
+        + r"\("
+        + r"\]"
+        + r"\["
+        + r"\}"
+        + r"\{"
+        + r"\|"
+        + "\\"
+        + r"\/"
+        + r"\*"
+        + r"]{1,}"
+    )  # noqa
+
+    _optional_components = ["tokenizer", "text_encoder"]
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKL,
+        transformer: Transformer3DModel,
+        scheduler: DPMSolverMultistepScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+    # Adapted from https://github.com/Open-Sora/Open-Sora/blob/master/diffusion/model/utils.py
+    def mask_text_embeddings(self, emb, mask):
+        if emb.shape[0] == 1:
+            keep_index = mask.sum().item()
+            return emb[:, :, :keep_index, :], keep_index
+        else:
+            masked_feature = emb * mask[:, None, :, None]
+            return masked_feature, emb.shape[2]
+
+    # Adapted from diffusers.pipelines.deepfloyd_if.pipeline_if.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: str = "",
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        negative_prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        clean_caption: bool = False,
+        max_sequence_length: int = 120,
+        **kwargs,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds`
+                instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For
+                Open-Sora, this should be "".
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                whether to use classifier free guidance or not
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                number of images that should be generated per prompt
+            device: (`torch.device`, *optional*):
+                torch device to place the resulting embeddings on
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For Open-Sora, it's should be the embeddings of the ""
+                string.
+            clean_caption (`bool`, defaults to `False`):
+                If `True`, the function will preprocess and clean the provided caption before encoding.
+            max_sequence_length (`int`, defaults to 120): Maximum sequence length to use for the prompt.
+        """
+
+        if "mask_feature" in kwargs:
+            deprecation_message = "The use of `mask_feature` is deprecated. It is no longer used in any computation and that doesn't affect the end results. It will be removed in a future version."
+            deprecate("mask_feature", "1.0.0", deprecation_message, standard_warn=False)
+
+        if device is None:
+            device = self._execution_device
+
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # See Section 3.1. of the paper.
+        max_length = max_sequence_length
+
+        if prompt_embeds is None:
+            prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+            text_inputs = self.tokenizer(
+                prompt,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                add_special_tokens=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                text_input_ids, untruncated_ids
+            ):
+                removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
+                logger.warning(
+                    "The following part of your input was truncated because CLIP can only handle sequences up to"
+                    f" {max_length} tokens: {removed_text}"
+                )
+
+            prompt_attention_mask = text_inputs.attention_mask
+            prompt_attention_mask = prompt_attention_mask.to(device)
+
+            prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)
+            prompt_embeds = prompt_embeds[0]
+
+        if self.text_encoder is not None:
+            dtype = self.text_encoder.dtype
+        elif self.transformer is not None:
+            dtype = self.transformer.dtype
+        else:
+            dtype = None
+
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+        prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            uncond_tokens = [negative_prompt] * batch_size
+            uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
+            max_length = prompt_embeds.shape[1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_attention_mask=True,
+                add_special_tokens=True,
+                return_tensors="pt",
+            )
+            negative_prompt_attention_mask = uncond_input.attention_mask
+            negative_prompt_attention_mask = negative_prompt_attention_mask.to(device)
+
+            negative_prompt_embeds = self.text_encoder(
+                uncond_input.input_ids.to(device), attention_mask=negative_prompt_attention_mask
+            )
+            negative_prompt_embeds = negative_prompt_embeds[0]
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1)
+            negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1)
+        else:
+            negative_prompt_embeds = None
+            negative_prompt_attention_mask = None
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    # 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,
+        prompt,
+        height,
+        width,
+        negative_prompt,
+        callback_steps,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        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)}."
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (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 prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
+    def _text_preprocessing(self, text, clean_caption=False):
+        if clean_caption and not is_bs4_available():
+            logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+            logger.warning("Setting `clean_caption` to False...")
+            clean_caption = False
+
+        if clean_caption and not is_ftfy_available():
+            logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+            logger.warning("Setting `clean_caption` to False...")
+            clean_caption = False
+
+        if not isinstance(text, (tuple, list)):
+            text = [text]
+
+        def process(text: str):
+            if clean_caption:
+                text = self._clean_caption(text)
+                text = self._clean_caption(text)
+            else:
+                text = text.lower().strip()
+            return text
+
+        return [process(t) for t in text]
+
+    # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
+    def _clean_caption(self, caption):
+        caption = str(caption)
+        caption = ul.unquote_plus(caption)
+        caption = caption.strip().lower()
+        caption = re.sub("<person>", "person", caption)
+        # urls:
+        caption = re.sub(
+            r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))",  # noqa
+            "",
+            caption,
+        )  # regex for urls
+        caption = re.sub(
+            r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))",  # noqa
+            "",
+            caption,
+        )  # regex for urls
+        # html:
+        caption = BeautifulSoup(caption, features="html.parser").text
+
+        # @<nickname>
+        caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+        # 31C0—31EF CJK Strokes
+        # 31F0—31FF Katakana Phonetic Extensions
+        # 3200—32FF Enclosed CJK Letters and Months
+        # 3300—33FF CJK Compatibility
+        # 3400—4DBF CJK Unified Ideographs Extension A
+        # 4DC0—4DFF Yijing Hexagram Symbols
+        # 4E00—9FFF CJK Unified Ideographs
+        caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+        caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+        caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+        caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+        caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+        caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+        caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+        #######################################################
+
+        # все виды тире / all types of dash --> "-"
+        caption = re.sub(
+            r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+",  # noqa
+            "-",
+            caption,
+        )
+
+        # кавычки к одному стандарту
+        caption = re.sub(r"[`´«»“”¨]", '"', caption)
+        caption = re.sub(r"[‘’]", "'", caption)
+
+        # &quot;
+        caption = re.sub(r"&quot;?", "", caption)
+        # &amp
+        caption = re.sub(r"&amp", "", caption)
+
+        # ip adresses:
+        caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+        # article ids:
+        caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+        # \n
+        caption = re.sub(r"\\n", " ", caption)
+
+        # "#123"
+        caption = re.sub(r"#\d{1,3}\b", "", caption)
+        # "#12345.."
+        caption = re.sub(r"#\d{5,}\b", "", caption)
+        # "123456.."
+        caption = re.sub(r"\b\d{6,}\b", "", caption)
+        # filenames:
+        caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+        #
+        caption = re.sub(r"[\"\']{2,}", r'"', caption)  # """AUSVERKAUFT"""
+        caption = re.sub(r"[\.]{2,}", r" ", caption)  # """AUSVERKAUFT"""
+
+        caption = re.sub(self.bad_punct_regex, r" ", caption)  # ***AUSVERKAUFT***, #AUSVERKAUFT
+        caption = re.sub(r"\s+\.\s+", r" ", caption)  # " . "
+
+        # this-is-my-cute-cat / this_is_my_cute_cat
+        regex2 = re.compile(r"(?:\-|\_)")
+        if len(re.findall(regex2, caption)) > 3:
+            caption = re.sub(regex2, " ", caption)
+
+        caption = ftfy.fix_text(caption)
+        caption = html.unescape(html.unescape(caption))
+
+        caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption)  # jc6640
+        caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption)  # jc6640vc
+        caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption)  # 6640vc231
+
+        caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+        caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+        caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+        caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+        caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+        caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption)  # j2d1a2a...
+
+        caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+        caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+        caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+        caption = re.sub(r"\s+", " ", caption)
+
+        caption.strip()
+
+        caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+        caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+        caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+        caption = re.sub(r"^\.\S+$", "", caption)
+
+        return caption.strip()
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_frames: int,
+        num_channels_latents: int,
+        height: int,
+        width: int,
+        dtype: torch.dtype,
+        device: Union[str, torch.device],
+        generator: torch.Generator,
+        latents: Optional[torch.FloatTensor] = None,
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_frames,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            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 decode_latents(self, latents: torch.FloatTensor, num_frames: int, decode_chunk_size: int = 14):
+        # [batch, channels, frames, height, width] -> [batch, frames, channels, height, width]
+        latents = latents.permute(0, 2, 1, 3, 4)
+        # [batch, frames, channels, height, width] -> [batch*frames, channels, height, width]
+        latents = latents.flatten(0, 1)
+
+        latents = 1 / self.vae.config.scaling_factor * latents
+
+        forward_vae_fn = self.vae._orig_mod.forward if is_compiled_module(self.vae) else self.vae.forward
+        accepts_num_frames = "num_frames" in set(inspect.signature(forward_vae_fn).parameters.keys())
+
+        # decode decode_chunk_size frames at a time to avoid OOM
+        frames = []
+        for i in range(0, latents.shape[0], decode_chunk_size):
+            num_frames_in = latents[i : i + decode_chunk_size].shape[0]
+            decode_kwargs = {}
+            if accepts_num_frames:
+                # we only pass num_frames_in if it's expected
+                decode_kwargs["num_frames"] = num_frames_in
+
+            frame = self.vae.decode(latents[i : i + decode_chunk_size], **decode_kwargs).sample
+            frames.append(frame)
+        frames = torch.cat(frames, dim=0)
+
+        # [batch*frames, channels, height, width] -> [batch, channels, frames, height, width]
+        frames = frames.reshape(-1, num_frames, *frames.shape[1:]).permute(0, 2, 1, 3, 4)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        frames = frames.float()
+        return frames
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: str = "",
+        num_inference_steps: int = 20,
+        timesteps: List[int] = None,
+        guidance_scale: float = 4.5,
+        min_guidance_scale: Optional[float] = None,
+        max_guidance_scale: Optional[float] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        clean_caption: bool = True,
+        max_sequence_length: int = 120,
+        decode_chunk_size: int = 4,
+        **kwargs,
+    ) -> Union[ImagePipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            num_inference_steps (`int`, *optional*, defaults to 100):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+                timesteps are used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 4.5):
+                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.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The width in pixels of the generated image.
+            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` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](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`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.FloatTensor`, *optional*): Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Alpha this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            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.IFPipelineOutput`] 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.
+            clean_caption (`bool`, *optional*, defaults to `True`):
+                Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+                be installed. If the dependencies are not installed, the embeddings will be created from the raw
+                prompt.
+            use_resolution_binning (`bool` defaults to `True`):
+                If set to `True`, the requested height and width are first mapped to the closest resolutions using
+                `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to
+                the requested resolution. Useful for generating non-square images.
+            max_sequence_length (`int` defaults to 120): Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ImagePipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is a list with the generated images
+        """
+        # 1. Check inputs. Raise error if not correct
+        num_frames = self.transformer.config.sample_size[0]
+        height = height or self.transformer.config.sample_size[1] * self.vae_scale_factor
+        width = width or self.transformer.config.sample_size[2] * self.vae_scale_factor
+
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            callback_steps,
+            prompt_embeds,
+            negative_prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_attention_mask,
+        )
+
+        # 2. Default height and width to transformer
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        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
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt,
+            do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            device=device,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            clean_caption=clean_caption,
+            max_sequence_length=max_sequence_length,
+        )
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+        # 5. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_frames,
+            latent_channels,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        # 8. Prepare guidance scale
+        # TODO: Hacky for testing, make this cleaner
+        if min_guidance_scale is not None and max_guidance_scale is not None:
+            guidance_scale = torch.linspace(min_guidance_scale, max_guidance_scale, num_frames).unsqueeze(0)
+            guidance_scale = guidance_scale.to(device, latents.dtype)
+            guidance_scale = guidance_scale.repeat(batch_size * num_images_per_prompt, 1)
+            guidance_scale = _append_dims(guidance_scale, latents.ndim).transpose(1, 2)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                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)
+
+                current_timestep = t
+                if not torch.is_tensor(current_timestep):
+                    # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+                    # This would be a good case for the `match` statement (Python 3.10+)
+                    is_mps = latent_model_input.device.type == "mps"
+                    if isinstance(current_timestep, float):
+                        dtype = torch.float32 if is_mps else torch.float64
+                    else:
+                        dtype = torch.int32 if is_mps else torch.int64
+                    current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device)
+                elif len(current_timestep.shape) == 0:
+                    current_timestep = current_timestep[None].to(latent_model_input.device)
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                current_timestep = current_timestep.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_attention_mask=prompt_attention_mask,
+                    timestep=current_timestep,
+                    return_dict=False,
+                )[0]
+
+                # learned sigma
+                if self.transformer.config.out_channels // 2 == latent_channels:
+                    noise_pred = noise_pred.chunk(2, dim=1)[0]
+                else:
+                    noise_pred = noise_pred
+
+                # 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 previous image: x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        if not output_type == "latent":
+            frames = self.decode_latents(latents, num_frames, decode_chunk_size)
+            frames = tensor2vid(frames, self.image_processor, output_type=output_type)
+        else:
+            frames = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (frames,)
+
+        return OpenSoraPipelineOutput(frames=frames)

src/diffusers/utils/dummy_pt_objects.py

@@ -227,6 +227,21 @@ class Transformer2DModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class Transformer3DModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class UNet1DModel(metaclass=DummyObject):
     _backends = ["torch"]
 

test.py

@@ -0,0 +1,39 @@
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+from diffusers import Transformer3DModel, AutoencoderTiny, DPMSolverMultistepScheduler, OpenSoraPipeline
+
+channels, num_frames, height, width, text_dim = 4, 2, 4, 4, 32
+
+model = Transformer3DModel(
+    in_channels=channels,
+    out_channels=channels*2,
+    cross_attention_dim=1408,
+    caption_channels=text_dim,
+    num_embeds_ada_norm=1000,
+    sample_size=(num_frames, height, width),
+)
+
+x = torch.randn(1, channels, num_frames, height, width)
+y = torch.randn(1, 77, 32)
+t = torch.ones(1)
+
+# with torch.no_grad():
+#     out = model(x, y, t)
+#     print(out.sample.shape)  # torch.Size([1, 8, 2, 4, 4])
+
+
+text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+tokenizer = T5Tokenizer.from_pretrained("google-t5/t5-small")
+vae = AutoencoderTiny.from_pretrained("madebyollin/taesd")
+scheduler = DPMSolverMultistepScheduler.from_pretrained("PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="scheduler")
+
+pipe = OpenSoraPipeline(
+    tokenizer=tokenizer,
+    text_encoder=text_encoder,
+    vae=vae,
+    transformer=model,
+    scheduler=scheduler,
+)
+
+prompt = ""
+out = pipe(prompt, num_inference_steps=1, min_guidance_scale=1.0, max_guidance_scale=3.0)