safety checker

scripts/convert_cosmos_to_diffusers.py

@@ -7,7 +7,16 @@ from accelerate import init_empty_weights
 from huggingface_hub import snapshot_download
 from transformers import T5EncoderModel, T5TokenizerFast
 
-from diffusers import AutoencoderKLCosmos, CosmosTextToWorldPipeline, CosmosTransformer3DModel, EDMEulerScheduler
+from diffusers import (
+    AutoencoderKLCosmos,
+    AutoencoderKLWan,
+    Cosmos2TextToImagePipeline,
+    Cosmos2VideoToWorldPipeline,
+    CosmosTextToWorldPipeline,
+    CosmosTransformer3DModel,
+    CosmosVideoToWorldPipeline,
+    EDMEulerScheduler,
+)
 
 
 def remove_keys_(key: str, state_dict: Dict[str, Any]):
@@ -29,7 +38,7 @@ def rename_transformer_blocks_(key: str, state_dict: Dict[str, Any]):
     state_dict[new_key] = state_dict.pop(key)
 
 
-TRANSFORMER_KEYS_RENAME_DICT = {
+TRANSFORMER_KEYS_RENAME_DICT_COSMOS_1_0 = {
     "t_embedder.1": "time_embed.t_embedder",
     "affline_norm": "time_embed.norm",
     ".blocks.0.block.attn": ".attn1",
@@ -56,7 +65,7 @@ TRANSFORMER_KEYS_RENAME_DICT = {
     "final_layer.linear": "proj_out",
 }
 
-TRANSFORMER_SPECIAL_KEYS_REMAP = {
+TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_1_0 = {
     "blocks.block": rename_transformer_blocks_,
     "logvar.0.freqs": remove_keys_,
     "logvar.0.phases": remove_keys_,
@@ -64,6 +73,45 @@ TRANSFORMER_SPECIAL_KEYS_REMAP = {
     "pos_embedder.seq": remove_keys_,
 }
 
+TRANSFORMER_KEYS_RENAME_DICT_COSMOS_2_0 = {
+    "t_embedder.1": "time_embed.t_embedder",
+    "t_embedding_norm": "time_embed.norm",
+    "blocks": "transformer_blocks",
+    "adaln_modulation_self_attn.1": "norm1.linear_1",
+    "adaln_modulation_self_attn.2": "norm1.linear_2",
+    "adaln_modulation_cross_attn.1": "norm2.linear_1",
+    "adaln_modulation_cross_attn.2": "norm2.linear_2",
+    "adaln_modulation_mlp.1": "norm3.linear_1",
+    "adaln_modulation_mlp.2": "norm3.linear_2",
+    "self_attn": "attn1",
+    "cross_attn": "attn2",
+    "q_proj": "to_q",
+    "k_proj": "to_k",
+    "v_proj": "to_v",
+    "output_proj": "to_out.0",
+    "q_norm": "norm_q",
+    "k_norm": "norm_k",
+    "mlp.layer1": "ff.net.0.proj",
+    "mlp.layer2": "ff.net.2",
+    "x_embedder.proj.1": "patch_embed.proj",
+    # "extra_pos_embedder": "learnable_pos_embed",
+    "final_layer.adaln_modulation.1": "norm_out.linear_1",
+    "final_layer.adaln_modulation.2": "norm_out.linear_2",
+    "final_layer.linear": "proj_out",
+}
+
+TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_2_0 = {
+    "accum_video_sample_counter": remove_keys_,
+    "accum_image_sample_counter": remove_keys_,
+    "accum_iteration": remove_keys_,
+    "accum_train_in_hours": remove_keys_,
+    "pos_embedder.seq": remove_keys_,
+    "pos_embedder.dim_spatial_range": remove_keys_,
+    "pos_embedder.dim_temporal_range": remove_keys_,
+    "_extra_state": remove_keys_,
+}
+
+
 TRANSFORMER_CONFIGS = {
     "Cosmos-1.0-Diffusion-7B-Text2World": {
         "in_channels": 16,
@@ -125,6 +173,66 @@ TRANSFORMER_CONFIGS = {
         "concat_padding_mask": True,
         "extra_pos_embed_type": "learnable",
     },
+    "Cosmos-2.0-Diffusion-2B-Text2Image": {
+        "in_channels": 16,
+        "out_channels": 16,
+        "num_attention_heads": 16,
+        "attention_head_dim": 128,
+        "num_layers": 28,
+        "mlp_ratio": 4.0,
+        "text_embed_dim": 1024,
+        "adaln_lora_dim": 256,
+        "max_size": (128, 240, 240),
+        "patch_size": (1, 2, 2),
+        "rope_scale": (1.0, 4.0, 4.0),
+        "concat_padding_mask": True,
+        "extra_pos_embed_type": None,
+    },
+    "Cosmos-2.0-Diffusion-14B-Text2Image": {
+        "in_channels": 16,
+        "out_channels": 16,
+        "num_attention_heads": 40,
+        "attention_head_dim": 128,
+        "num_layers": 36,
+        "mlp_ratio": 4.0,
+        "text_embed_dim": 1024,
+        "adaln_lora_dim": 256,
+        "max_size": (128, 240, 240),
+        "patch_size": (1, 2, 2),
+        "rope_scale": (1.0, 4.0, 4.0),
+        "concat_padding_mask": True,
+        "extra_pos_embed_type": None,
+    },
+    "Cosmos-2.0-Diffusion-2B-Video2World": {
+        "in_channels": 16 + 1,
+        "out_channels": 16,
+        "num_attention_heads": 16,
+        "attention_head_dim": 128,
+        "num_layers": 28,
+        "mlp_ratio": 4.0,
+        "text_embed_dim": 1024,
+        "adaln_lora_dim": 256,
+        "max_size": (128, 240, 240),
+        "patch_size": (1, 2, 2),
+        "rope_scale": (1.0, 3.0, 3.0),
+        "concat_padding_mask": True,
+        "extra_pos_embed_type": None,
+    },
+    "Cosmos-2.0-Diffusion-14B-Video2World": {
+        "in_channels": 16 + 1,
+        "out_channels": 16,
+        "num_attention_heads": 40,
+        "attention_head_dim": 128,
+        "num_layers": 36,
+        "mlp_ratio": 4.0,
+        "text_embed_dim": 1024,
+        "adaln_lora_dim": 256,
+        "max_size": (128, 240, 240),
+        "patch_size": (1, 2, 2),
+        "rope_scale": (20 / 24, 2.0, 2.0),
+        "concat_padding_mask": True,
+        "extra_pos_embed_type": None,
+    },
 }
 
 VAE_KEYS_RENAME_DICT = {
@@ -216,9 +324,18 @@ def get_state_dict(saved_dict: Dict[str, Any]) -> Dict[str, Any]:
     return state_dict
 
 
-def convert_transformer(transformer_type: str, ckpt_path: str):
+def convert_transformer(transformer_type: str, ckpt_path: str, weights_only: bool = True):
     PREFIX_KEY = "net."
-    original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", weights_only=True))
+    original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", weights_only=weights_only))
+
+    if "Cosmos-1.0" in transformer_type:
+        TRANSFORMER_KEYS_RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT_COSMOS_1_0
+        TRANSFORMER_SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_1_0
+    elif "Cosmos-2.0" in transformer_type:
+        TRANSFORMER_KEYS_RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT_COSMOS_2_0
+        TRANSFORMER_SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_2_0
+    else:
+        assert False
 
     with init_empty_weights():
         config = TRANSFORMER_CONFIGS[transformer_type]
@@ -281,13 +398,69 @@ def convert_vae(vae_type: str):
     return vae
 
 
+def save_pipeline_cosmos_1_0(args, transformer, vae):
+    text_encoder = T5EncoderModel.from_pretrained(args.text_encoder_path, torch_dtype=torch.bfloat16)
+    tokenizer = T5TokenizerFast.from_pretrained(args.tokenizer_path)
+    # The original code initializes EDM config with sigma_min=0.0002, but does not make use of it anywhere directly.
+    # So, the sigma_min values that is used is the default value of 0.002.
+    scheduler = EDMEulerScheduler(
+        sigma_min=0.002,
+        sigma_max=80,
+        sigma_data=0.5,
+        sigma_schedule="karras",
+        num_train_timesteps=1000,
+        prediction_type="epsilon",
+        rho=7.0,
+        final_sigmas_type="sigma_min",
+    )
+
+    pipe_cls = CosmosTextToWorldPipeline if "Text2World" in args.transformer_type else CosmosVideoToWorldPipeline
+    pipe = pipe_cls(
+        text_encoder=text_encoder,
+        tokenizer=tokenizer,
+        transformer=transformer,
+        vae=vae,
+        scheduler=scheduler,
+    )
+    pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")
+
+
+def save_pipeline_cosmos_2_0(args, transformer, vae):
+    text_encoder = T5EncoderModel.from_pretrained(args.text_encoder_path, torch_dtype=torch.bfloat16)
+    tokenizer = T5TokenizerFast.from_pretrained(args.tokenizer_path)
+
+    scheduler = EDMEulerScheduler(
+        sigma_min=0.002,
+        sigma_max=80,
+        sigma_data=1.0,
+        sigma_schedule="karras",
+        num_train_timesteps=1000,
+        prediction_type="epsilon",
+        rho=7.0,
+        final_sigmas_type="sigma_min",
+        use_flow_sigmas=True,
+    )
+
+    pipe_cls = Cosmos2TextToImagePipeline if "Text2Image" in args.transformer_type else Cosmos2VideoToWorldPipeline
+    pipe = pipe_cls(
+        text_encoder=text_encoder,
+        tokenizer=tokenizer,
+        transformer=transformer,
+        vae=vae,
+        scheduler=scheduler,
+    )
+    pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")
+
+
 def get_args():
     parser = argparse.ArgumentParser()
     parser.add_argument("--transformer_type", type=str, default=None, choices=list(TRANSFORMER_CONFIGS.keys()))
     parser.add_argument(
         "--transformer_ckpt_path", type=str, default=None, help="Path to original transformer checkpoint"
     )
-    parser.add_argument("--vae_type", type=str, default=None, choices=list(VAE_CONFIGS.keys()), help="Type of VAE")
+    parser.add_argument(
+        "--vae_type", type=str, default=None, choices=["none", *list(VAE_CONFIGS.keys())], help="Type of VAE"
+    )
     parser.add_argument("--text_encoder_path", type=str, default="google-t5/t5-11b")
     parser.add_argument("--tokenizer_path", type=str, default="google-t5/t5-11b")
     parser.add_argument("--save_pipeline", action="store_true")
@@ -316,37 +489,26 @@ if __name__ == "__main__":
         assert args.tokenizer_path is not None
 
     if args.transformer_ckpt_path is not None:
-        transformer = convert_transformer(args.transformer_type, args.transformer_ckpt_path)
+        weights_only = "Cosmos-1.0" in args.transformer_type
+        transformer = convert_transformer(args.transformer_type, args.transformer_ckpt_path, weights_only)
         transformer = transformer.to(dtype=dtype)
         if not args.save_pipeline:
             transformer.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")
 
     if args.vae_type is not None:
-        vae = convert_vae(args.vae_type)
+        if "Cosmos-1.0" in args.transformer_type:
+            vae = convert_vae(args.vae_type)
+        else:
+            vae = AutoencoderKLWan.from_pretrained(
+                "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", subfolder="vae", torch_dtype=torch.float32
+            )
         if not args.save_pipeline:
             vae.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")
 
     if args.save_pipeline:
-        text_encoder = T5EncoderModel.from_pretrained(args.text_encoder_path, torch_dtype=dtype)
-        tokenizer = T5TokenizerFast.from_pretrained(args.tokenizer_path)
-        # The original code initializes EDM config with sigma_min=0.0002, but does not make use of it anywhere directly.
-        # So, the sigma_min values that is used is the default value of 0.002.
-        scheduler = EDMEulerScheduler(
-            sigma_min=0.002,
-            sigma_max=80,
-            sigma_data=0.5,
-            sigma_schedule="karras",
-            num_train_timesteps=1000,
-            prediction_type="epsilon",
-            rho=7.0,
-            final_sigmas_type="sigma_min",
-        )
-
-        pipe = CosmosTextToWorldPipeline(
-            text_encoder=text_encoder,
-            tokenizer=tokenizer,
-            transformer=transformer,
-            vae=vae,
-            scheduler=scheduler,
-        )
-        pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")
+        if "Cosmos-1.0" in args.transformer_type:
+            save_pipeline_cosmos_1_0(args, transformer, vae)
+        elif "Cosmos-2.0" in args.transformer_type:
+            save_pipeline_cosmos_2_0(args, transformer, vae)
+        else:
+            assert False

src/diffusers/__init__.py

@@ -361,6 +361,8 @@ else:
             "CogView4ControlPipeline",
             "CogView4Pipeline",
             "ConsisIDPipeline",
+            "Cosmos2TextToImagePipeline",
+            "Cosmos2VideoToWorldPipeline",
             "CosmosTextToWorldPipeline",
             "CosmosVideoToWorldPipeline",
             "CycleDiffusionPipeline",
@@ -949,6 +951,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             CogView4ControlPipeline,
             CogView4Pipeline,
             ConsisIDPipeline,
+            Cosmos2TextToImagePipeline,
+            Cosmos2VideoToWorldPipeline,
             CosmosTextToWorldPipeline,
             CosmosVideoToWorldPipeline,
             CycleDiffusionPipeline,

src/diffusers/models/transformers/transformer_cosmos.py

@@ -100,11 +100,15 @@ class CosmosAdaLayerNorm(nn.Module):
         embedded_timestep = self.linear_2(embedded_timestep)
 
         if temb is not None:
-            embedded_timestep = embedded_timestep + temb[:, : 2 * self.embedding_dim]
+            embedded_timestep = embedded_timestep + temb[..., : 2 * self.embedding_dim]
 
-        shift, scale = embedded_timestep.chunk(2, dim=1)
+        shift, scale = embedded_timestep.chunk(2, dim=-1)
         hidden_states = self.norm(hidden_states)
-        hidden_states = hidden_states * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+        if embedded_timestep.ndim == 2:
+            shift, scale = (x.unsqueeze(1) for x in (shift, scale))
+
+        hidden_states = hidden_states * (1 + scale) + shift
         return hidden_states
 
 
@@ -135,9 +139,13 @@ class CosmosAdaLayerNormZero(nn.Module):
         if temb is not None:
             embedded_timestep = embedded_timestep + temb
 
-        shift, scale, gate = embedded_timestep.chunk(3, dim=1)
+        shift, scale, gate = embedded_timestep.chunk(3, dim=-1)
         hidden_states = self.norm(hidden_states)
-        hidden_states = hidden_states * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+        if embedded_timestep.ndim == 2:
+            shift, scale, gate = (x.unsqueeze(1) for x in (shift, scale, gate))
+
+        hidden_states = hidden_states * (1 + scale) + shift
         return hidden_states, gate
 
 
@@ -255,19 +263,19 @@ class CosmosTransformerBlock(nn.Module):
         # 1. Self Attention
         norm_hidden_states, gate = self.norm1(hidden_states, embedded_timestep, temb)
         attn_output = self.attn1(norm_hidden_states, image_rotary_emb=image_rotary_emb)
-        hidden_states = hidden_states + gate.unsqueeze(1) * attn_output
+        hidden_states = hidden_states + gate * attn_output
 
         # 2. Cross Attention
         norm_hidden_states, gate = self.norm2(hidden_states, embedded_timestep, temb)
         attn_output = self.attn2(
             norm_hidden_states, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask
         )
-        hidden_states = hidden_states + gate.unsqueeze(1) * attn_output
+        hidden_states = hidden_states + gate * attn_output
 
         # 3. Feed Forward
         norm_hidden_states, gate = self.norm3(hidden_states, embedded_timestep, temb)
         ff_output = self.ff(norm_hidden_states)
-        hidden_states = hidden_states + gate.unsqueeze(1) * ff_output
+        hidden_states = hidden_states + gate * ff_output
 
         return hidden_states
 
@@ -513,7 +521,23 @@ class CosmosTransformer3DModel(ModelMixin, ConfigMixin):
         hidden_states = hidden_states.flatten(1, 3)  # [B, T, H, W, C] -> [B, THW, C]
 
         # 4. Timestep embeddings
-        temb, embedded_timestep = self.time_embed(hidden_states, timestep)
+        if timestep.ndim == 1:
+            temb, embedded_timestep = self.time_embed(hidden_states, timestep)
+        elif timestep.ndim == 5:
+            assert timestep.shape == (batch_size, 1, num_frames, 1, 1), (
+                f"Expected timestep to have shape [B, 1, T, 1, 1], but got {timestep.shape}"
+            )
+            timestep = timestep.flatten()
+            temb, embedded_timestep = self.time_embed(hidden_states, timestep)
+            # We can do this because num_frames == post_patch_num_frames, as p_t is 1
+            temb, embedded_timestep = (
+                x.view(batch_size, post_patch_num_frames, 1, 1, -1)
+                .expand(-1, -1, post_patch_height, post_patch_width, -1)
+                .flatten(1, 3)
+                for x in (temb, embedded_timestep)
+            )  # [BT, C] -> [B, T, 1, 1, C] -> [B, T, H, W, C] -> [B, THW, C]
+        else:
+            assert False
 
         # 5. Transformer blocks
         for block in self.transformer_blocks:
@@ -544,8 +568,6 @@ class CosmosTransformer3DModel(ModelMixin, ConfigMixin):
         hidden_states = self.proj_out(hidden_states)
         hidden_states = hidden_states.unflatten(2, (p_h, p_w, p_t, -1))
         hidden_states = hidden_states.unflatten(1, (post_patch_num_frames, post_patch_height, post_patch_width))
-        # Please just kill me at this point. What even is this permutation order and why is it different from the patching order?
-        # Another few hours of sanity lost to the void.
         hidden_states = hidden_states.permute(0, 7, 1, 6, 2, 4, 3, 5)
         hidden_states = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
 

src/diffusers/pipelines/__init__.py

@@ -157,7 +157,12 @@ else:
     _import_structure["cogview3"] = ["CogView3PlusPipeline"]
     _import_structure["cogview4"] = ["CogView4Pipeline", "CogView4ControlPipeline"]
     _import_structure["consisid"] = ["ConsisIDPipeline"]
-    _import_structure["cosmos"] = ["CosmosTextToWorldPipeline", "CosmosVideoToWorldPipeline"]
+    _import_structure["cosmos"] = [
+        "Cosmos2TextToImagePipeline",
+        "CosmosTextToWorldPipeline",
+        "CosmosVideoToWorldPipeline",
+        "Cosmos2VideoToWorldPipeline",
+    ]
     _import_structure["controlnet"].extend(
         [
             "BlipDiffusionControlNetPipeline",
@@ -559,7 +564,12 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionControlNetXSPipeline,
             StableDiffusionXLControlNetXSPipeline,
         )
-        from .cosmos import CosmosTextToWorldPipeline, CosmosVideoToWorldPipeline
+        from .cosmos import (
+            Cosmos2TextToImagePipeline,
+            Cosmos2VideoToWorldPipeline,
+            CosmosTextToWorldPipeline,
+            CosmosVideoToWorldPipeline,
+        )
         from .deepfloyd_if import (
             IFImg2ImgPipeline,
             IFImg2ImgSuperResolutionPipeline,

src/diffusers/pipelines/cosmos/__init__.py

@@ -22,6 +22,8 @@ except OptionalDependencyNotAvailable:
 
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
+    _import_structure["pipeline_cosmos2_text2image"] = ["Cosmos2TextToImagePipeline"]
+    _import_structure["pipeline_cosmos2_video2world"] = ["Cosmos2VideoToWorldPipeline"]
     _import_structure["pipeline_cosmos_text2world"] = ["CosmosTextToWorldPipeline"]
     _import_structure["pipeline_cosmos_video2world"] = ["CosmosVideoToWorldPipeline"]
 
@@ -33,6 +35,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     except OptionalDependencyNotAvailable:
         from ...utils.dummy_torch_and_transformers_objects import *
     else:
+        from .pipeline_cosmos2_text2image import Cosmos2TextToImagePipeline
+        from .pipeline_cosmos2_video2world import Cosmos2VideoToWorldPipeline
         from .pipeline_cosmos_text2world import CosmosTextToWorldPipeline
         from .pipeline_cosmos_video2world import CosmosVideoToWorldPipeline
 

src/diffusers/pipelines/cosmos/pipeline_cosmos2_text2image.py

@@ -0,0 +1,653 @@
+# Copyright 2025 The NVIDIA Team 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 inspect
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...models import AutoencoderKLWan, CosmosTransformer3DModel
+from ...schedulers import EDMEulerScheduler
+from ...utils import is_cosmos_guardrail_available, is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import CosmosImagePipelineOutput
+
+
+if is_cosmos_guardrail_available():
+    from cosmos_guardrail import CosmosSafetyChecker
+else:
+
+    class CosmosSafetyChecker:
+        def __init__(self, *args, **kwargs):
+            raise ImportError(
+                "`cosmos_guardrail` is not installed. Please install it to use the safety checker for Cosmos: `pip install cosmos_guardrail`."
+            )
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers import Cosmos2TextToImagePipeline
+
+        >>> # Available checkpoints: nvidia/Cosmos-Predict2-2B-Text2Image, nvidia/Cosmos-Predict2-14B-Text2Image
+        >>> model_id = "nvidia/Cosmos-Predict2-2B-Text2Image"
+        >>> pipe = Cosmos2TextToImagePipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "A close-up shot captures a vibrant yellow scrubber vigorously working on a grimy plate, its bristles moving in circular motions to lift stubborn grease and food residue. The dish, once covered in remnants of a hearty meal, gradually reveals its original glossy surface. Suds form and bubble around the scrubber, creating a satisfying visual of cleanliness in progress. The sound of scrubbing fills the air, accompanied by the gentle clinking of the dish against the sink. As the scrubber continues its task, the dish transforms, gleaming under the bright kitchen lights, symbolizing the triumph of cleanliness over mess."
+        >>> negative_prompt = "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, jerky movements, low frame rate, artifacting, color banding, unnatural transitions, outdated special effects, fake elements, unconvincing visuals, poorly edited content, jump cuts, visual noise, and flickering. Overall, the video is of poor quality."
+
+        >>> output = pipe(
+        ...     prompt=prompt, negative_prompt=negative_prompt, generator=torch.Generator().manual_seed(1)
+        ... ).images[0]
+        >>> output.save("output.png")
+        ```
+"""
+
+
+# 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,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    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 override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` 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 and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    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)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, 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 Cosmos2TextToImagePipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using [Cosmos](https://github.com/NVIDIA/Cosmos).
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        text_encoder ([`T5EncoderModel`]):
+            Frozen text-encoder. Cosmos uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel); specifically the
+            [t5-11b](https://huggingface.co/google-t5/t5-11b) variant.
+        tokenizer (`T5TokenizerFast`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`CosmosTransformer3DModel`]):
+            Conditional Transformer to denoise the encoded image latents.
+        scheduler ([`EDMEulerScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLWan`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+    # We mark safety_checker as optional here to get around some test failures, but it is not really optional
+    _optional_components = ["safety_checker"]
+
+    def __init__(
+        self,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: CosmosTransformer3DModel,
+        vae: AutoencoderKLWan,
+        scheduler: EDMEulerScheduler,
+        safety_checker: CosmosSafetyChecker = None,
+    ):
+        super().__init__()
+
+        if safety_checker is None:
+            safety_checker = CosmosSafetyChecker()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+        )
+
+        self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    # Copied from diffusers.pipelines.cosmos.pipeline_cosmos_text2world.CosmosTextToWorldPipeline._get_t5_prompt_embeds
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        max_sequence_length: int = 512,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+            return_length=True,
+            return_offsets_mapping=False,
+        )
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask.bool().to(device)
+
+        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_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(
+            text_input_ids.to(device), attention_mask=prompt_attention_mask
+        ).last_hidden_state
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        lengths = prompt_attention_mask.sum(dim=1).cpu()
+        for i, length in enumerate(lengths):
+            prompt_embeds[i, length:] = 0
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.cosmos.pipeline_cosmos_text2world.CosmosTextToWorldPipeline.encode_prompt with num_videos_per_prompt->num_images_per_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_images_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 512,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        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 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`).
+            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 videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *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.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt, max_sequence_length=max_sequence_length, device=device, dtype=dtype
+            )
+
+            # duplicate text embeddings for each generation per prompt, using mps friendly method
+            _, seq_len, _ = prompt_embeds.shape
+            prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and 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 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`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt, max_sequence_length=max_sequence_length, device=device, dtype=dtype
+            )
+
+            # duplicate text embeddings for each generation per prompt, using mps friendly method
+            _, seq_len, _ = negative_prompt_embeds.shape
+            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)
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_channels_latents: 16,
+        height: int = 768,
+        width: int = 1360,
+        num_frames: int = 1,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype) * self.scheduler.config.sigma_max
+
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        latent_height = height // self.vae_scale_factor_spatial
+        latent_width = width // self.vae_scale_factor_spatial
+        shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width)
+
+        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."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents * self.scheduler.config.sigma_max
+
+    # Copied from diffusers.pipelines.cosmos.pipeline_cosmos_text2world.CosmosTextToWorldPipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if height % 16 != 0 or width % 16 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        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)}")
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 768,
+        width: int = 1360,
+        num_inference_steps: int = 35,
+        guidance_scale: float = 7.0,
+        num_images_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ):
+        r"""
+        The call function to 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.
+            height (`int`, defaults to `768`):
+                The height in pixels of the generated image.
+            width (`int`, defaults to `1360`):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, defaults to `35`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, defaults to `7.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
+                `guidance_scale > 1`.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *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 is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *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 PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`CosmosImagePipelineOutput`] instead of a plain tuple.
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~CosmosImagePipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`CosmosImagePipelineOutput`] is returned, otherwise a `tuple` is returned
+                where the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if self.safety_checker is None:
+            raise ValueError(
+                f"You have disabled the safety checker for {self.__class__}. This is in violation of the "
+                "[NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license). "
+                f"Please ensure that you are compliant with the license agreement."
+            )
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        num_frames = 1
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(prompt, height, width, prompt_embeds, callback_on_step_end_tensor_inputs)
+
+        self._guidance_scale = guidance_scale
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        if self.safety_checker is not None:
+            self.safety_checker.to(device)
+            if prompt is not None:
+                prompt_list = [prompt] if isinstance(prompt, str) else prompt
+                for p in prompt_list:
+                    if not self.safety_checker.check_text_safety(p):
+                        raise ValueError(
+                            f"Cosmos Guardrail detected unsafe text in the prompt: {p}. Please ensure that the "
+                            f"prompt abides by the NVIDIA Open Model License Agreement."
+                        )
+            self.safety_checker.to("cpu")
+
+        # 2. Define call parameters
+        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]
+
+        # 3. Encode input prompt
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_images_per_prompt=num_images_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            device=device,
+            max_sequence_length=max_sequence_length,
+        )
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device)
+
+        # 5. Prepare latent variables
+        transformer_dtype = self.transformer.dtype
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        padding_mask = latents.new_zeros(1, 1, height, width, dtype=transformer_dtype)
+
+        # 6. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                timestep = t.expand(latents.shape[0]).to(transformer_dtype)
+                current_sigma = self.scheduler.sigmas[i]
+
+                latent_model_input = latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+                latent_model_input = latent_model_input.to(transformer_dtype)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    padding_mask=padding_mask,
+                    return_dict=False,
+                )[0]
+                noise_pred = self.scheduler.precondition_outputs(latents, noise_pred, current_sigma)
+
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        padding_mask=padding_mask,
+                        return_dict=False,
+                    )[0]
+                    noise_pred_uncond = self.scheduler.precondition_outputs(latents, noise_pred_uncond, current_sigma)
+                    noise_pred = noise_pred + self.guidance_scale * (noise_pred - noise_pred_uncond)
+
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, pred_original_sample=noise_pred, return_dict=False
+                )[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # 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 XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std / self.scheduler.config.sigma_data + latents_mean
+            video = self.vae.decode(latents.to(self.vae.dtype), return_dict=False)[0]
+
+            if self.safety_checker is not None:
+                self.safety_checker.to(device)
+                video = self.video_processor.postprocess_video(video, output_type="np")
+                video = (video * 255).astype(np.uint8)
+                video_batch = []
+                for vid in video:
+                    vid = self.safety_checker.check_video_safety(vid)
+                    video_batch.append(vid)
+                video = np.stack(video_batch).astype(np.float32) / 255.0 * 2 - 1
+                video = torch.from_numpy(video).permute(0, 4, 1, 2, 3)
+                video = self.video_processor.postprocess_video(video, output_type=output_type)
+                self.safety_checker.to("cpu")
+            else:
+                video = self.video_processor.postprocess_video(video, output_type=output_type)
+            image = [batch[0] for batch in video]
+            if isinstance(video, torch.Tensor):
+                image = torch.stack(image)
+            elif isinstance(video, np.ndarray):
+                image = np.stack(image)
+        else:
+            image = latents[:, :, 0]
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return CosmosImagePipelineOutput(images=image)

src/diffusers/pipelines/cosmos/pipeline_cosmos2_video2world.py

@@ -0,0 +1,771 @@
+# Copyright 2025 The NVIDIA Team 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 inspect
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...image_processor import PipelineImageInput
+from ...models import AutoencoderKLWan, CosmosTransformer3DModel
+from ...schedulers import EDMEulerScheduler
+from ...utils import is_cosmos_guardrail_available, is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import CosmosPipelineOutput
+
+
+if is_cosmos_guardrail_available():
+    from cosmos_guardrail import CosmosSafetyChecker
+else:
+
+    class CosmosSafetyChecker:
+        def __init__(self, *args, **kwargs):
+            raise ImportError(
+                "`cosmos_guardrail` is not installed. Please install it to use the safety checker for Cosmos: `pip install cosmos_guardrail`."
+            )
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers import Cosmos2VideoToWorldPipeline
+        >>> from diffusers.utils import export_to_video, load_image
+
+        >>> # Available checkpoints: nvidia/Cosmos-Predict2-2B-Video2World, nvidia/Cosmos-Predict2-14B-Video2World
+        >>> model_id = "nvidia/Cosmos-Predict2-2B-Video2World"
+        >>> pipe = Cosmos2VideoToWorldPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "A close-up shot captures a vibrant yellow scrubber vigorously working on a grimy plate, its bristles moving in circular motions to lift stubborn grease and food residue. The dish, once covered in remnants of a hearty meal, gradually reveals its original glossy surface. Suds form and bubble around the scrubber, creating a satisfying visual of cleanliness in progress. The sound of scrubbing fills the air, accompanied by the gentle clinking of the dish against the sink. As the scrubber continues its task, the dish transforms, gleaming under the bright kitchen lights, symbolizing the triumph of cleanliness over mess."
+        >>> negative_prompt = "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, jerky movements, low frame rate, artifacting, color banding, unnatural transitions, outdated special effects, fake elements, unconvincing visuals, poorly edited content, jump cuts, visual noise, and flickering. Overall, the video is of poor quality."
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/yellow-scrubber.png"
+        ... )
+
+        >>> video = pipe(
+        ...     image=image, prompt=prompt, negative_prompt=negative_prompt, generator=torch.Generator().manual_seed(1)
+        ... ).frames[0]
+        >>> export_to_video(video, "output.mp4", fps=16)
+        ```
+"""
+
+
+# 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,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    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 override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` 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 and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    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)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, 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
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class Cosmos2VideoToWorldPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using [Cosmos](https://github.com/NVIDIA/Cosmos).
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        text_encoder ([`T5EncoderModel`]):
+            Frozen text-encoder. Cosmos uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel); specifically the
+            [t5-11b](https://huggingface.co/google-t5/t5-11b) variant.
+        tokenizer (`T5TokenizerFast`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`CosmosTransformer3DModel`]):
+            Conditional Transformer to denoise the encoded image latents.
+        scheduler ([`EDMEulerScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLWan`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+    # We mark safety_checker as optional here to get around some test failures, but it is not really optional
+    _optional_components = ["safety_checker"]
+
+    def __init__(
+        self,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: CosmosTransformer3DModel,
+        vae: AutoencoderKLWan,
+        scheduler: EDMEulerScheduler,
+        safety_checker: CosmosSafetyChecker = None,
+    ):
+        super().__init__()
+
+        if safety_checker is None:
+            safety_checker = CosmosSafetyChecker()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+        )
+
+        self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    # Copied from diffusers.pipelines.cosmos.pipeline_cosmos_text2world.CosmosTextToWorldPipeline._get_t5_prompt_embeds
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        max_sequence_length: int = 512,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+            return_length=True,
+            return_offsets_mapping=False,
+        )
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask.bool().to(device)
+
+        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_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(
+            text_input_ids.to(device), attention_mask=prompt_attention_mask
+        ).last_hidden_state
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        lengths = prompt_attention_mask.sum(dim=1).cpu()
+        for i, length in enumerate(lengths):
+            prompt_embeds[i, length:] = 0
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.cosmos.pipeline_cosmos_text2world.CosmosTextToWorldPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 512,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        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 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`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *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.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt, max_sequence_length=max_sequence_length, device=device, dtype=dtype
+            )
+
+            # duplicate text embeddings for each generation per prompt, using mps friendly method
+            _, seq_len, _ = prompt_embeds.shape
+            prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+            prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and 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 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`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt, max_sequence_length=max_sequence_length, device=device, dtype=dtype
+            )
+
+            # duplicate text embeddings for each generation per prompt, using mps friendly method
+            _, seq_len, _ = negative_prompt_embeds.shape
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def prepare_latents(
+        self,
+        video: torch.Tensor,
+        batch_size: int,
+        num_channels_latents: 16,
+        height: int = 704,
+        width: int = 1280,
+        num_frames: int = 77,
+        do_classifier_free_guidance: bool = True,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        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."
+            )
+
+        num_cond_frames = video.size(2)
+        if num_cond_frames >= num_frames:
+            # Take the last `num_frames` frames for conditioning
+            num_cond_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+            video = video[:, :, -num_frames:]
+        else:
+            num_cond_latent_frames = (num_cond_frames - 1) // self.vae_scale_factor_temporal + 1
+            num_padding_frames = num_frames - num_cond_frames
+            last_frame = video[:, :, -1:]
+            padding = last_frame.repeat(1, 1, num_padding_frames, 1, 1)
+            video = torch.cat([video, padding], dim=2)
+
+        if isinstance(generator, list):
+            init_latents = [
+                retrieve_latents(self.vae.encode(video[i].unsqueeze(0)), generator=generator[i])
+                for i in range(batch_size)
+            ]
+        else:
+            init_latents = [retrieve_latents(self.vae.encode(vid.unsqueeze(0)), generator) for vid in video]
+
+        init_latents = torch.cat(init_latents, dim=0).to(dtype)
+
+        latents_mean = (
+            torch.tensor(self.vae.config.latents_mean).view(1, self.vae.config.z_dim, 1, 1, 1).to(device, dtype)
+        )
+        latents_std = (
+            torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(device, dtype)
+        )
+        init_latents = (init_latents - latents_mean) / latents_std * self.scheduler.config.sigma_data
+
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        latent_height = height // self.vae_scale_factor_spatial
+        latent_width = width // self.vae_scale_factor_spatial
+        shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width)
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device=device, dtype=dtype)
+
+        latents = latents * self.scheduler.config.sigma_max
+
+        padding_shape = (batch_size, 1, num_latent_frames, latent_height, latent_width)
+        ones_padding = latents.new_ones(padding_shape)
+        zeros_padding = latents.new_zeros(padding_shape)
+
+        cond_indicator = latents.new_zeros(1, 1, latents.size(2), 1, 1)
+        cond_indicator[:, :, :num_cond_latent_frames] = 1.0
+        cond_mask = cond_indicator * ones_padding + (1 - cond_indicator) * zeros_padding
+
+        uncond_indicator = uncond_mask = None
+        if do_classifier_free_guidance:
+            uncond_indicator = latents.new_zeros(1, 1, latents.size(2), 1, 1)
+            uncond_indicator[:, :, :num_cond_latent_frames] = 1.0
+            uncond_mask = uncond_indicator * ones_padding + (1 - uncond_indicator) * zeros_padding
+
+        return latents, init_latents, cond_indicator, uncond_indicator, cond_mask, uncond_mask
+
+    # Copied from diffusers.pipelines.cosmos.pipeline_cosmos_text2world.CosmosTextToWorldPipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if height % 16 != 0 or width % 16 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        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)}")
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: PipelineImageInput = None,
+        video: List[PipelineImageInput] = None,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 704,
+        width: int = 1280,
+        num_frames: int = 77,
+        num_inference_steps: int = 35,
+        guidance_scale: float = 7.0,
+        fps: int = 16,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+        sigma_conditioning: float = 0.0001,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            image (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, *optional*):
+                The image to be used as a conditioning input for the video generation.
+            video (`List[PIL.Image.Image]`, `np.ndarray`, `torch.Tensor`, *optional*):
+                The video to be used as a conditioning input for the video generation.
+            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.
+            height (`int`, defaults to `704`):
+                The height in pixels of the generated image.
+            width (`int`, defaults to `1280`):
+                The width in pixels of the generated image.
+            num_frames (`int`, defaults to `77`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `35`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, defaults to `7.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
+                `guidance_scale > 1`.
+            fps (`int`, defaults to `16`):
+                The frames per second of the generated video.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *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 is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *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 PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`CosmosPipelineOutput`] instead of a plain tuple.
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int`, defaults to `512`):
+                The maximum number of tokens in the prompt. If the prompt exceeds this length, it will be truncated. If
+                the prompt is shorter than this length, it will be padded.
+            sigma_conditioning (`float`, defaults to `0.0001`):
+                The sigma value used for scaling conditioning latents. Ideally, it should not be changed or should be
+                set to a small value close to zero.
+
+        Examples:
+
+        Returns:
+            [`~CosmosPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`CosmosPipelineOutput`] is returned, otherwise a `tuple` is returned where
+                the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if self.safety_checker is None:
+            raise ValueError(
+                f"You have disabled the safety checker for {self.__class__}. This is in violation of the "
+                "[NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license). "
+                f"Please ensure that you are compliant with the license agreement."
+            )
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(prompt, height, width, prompt_embeds, callback_on_step_end_tensor_inputs)
+
+        self._guidance_scale = guidance_scale
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        if self.safety_checker is not None:
+            self.safety_checker.to(device)
+            if prompt is not None:
+                prompt_list = [prompt] if isinstance(prompt, str) else prompt
+                for p in prompt_list:
+                    if not self.safety_checker.check_text_safety(p):
+                        raise ValueError(
+                            f"Cosmos Guardrail detected unsafe text in the prompt: {p}. Please ensure that the "
+                            f"prompt abides by the NVIDIA Open Model License Agreement."
+                        )
+            self.safety_checker.to("cpu")
+
+        # 2. Define call parameters
+        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]
+
+        # 3. Encode input prompt
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            device=device,
+            max_sequence_length=max_sequence_length,
+        )
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device)
+
+        # 5. Prepare latent variables
+        vae_dtype = self.vae.dtype
+        transformer_dtype = self.transformer.dtype
+
+        if image is not None:
+            video = self.video_processor.preprocess(image, height, width).unsqueeze(2)
+        else:
+            video = self.video_processor.preprocess_video(video, height, width)
+        video = video.to(device=device, dtype=vae_dtype)
+
+        num_channels_latents = self.transformer.config.in_channels - 1
+        latents, conditioning_latents, cond_indicator, uncond_indicator, cond_mask, uncond_mask = self.prepare_latents(
+            video,
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            self.do_classifier_free_guidance,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+        unconditioning_latents = None
+
+        cond_mask = cond_mask.to(transformer_dtype)
+        if self.do_classifier_free_guidance:
+            uncond_mask = uncond_mask.to(transformer_dtype)
+            unconditioning_latents = conditioning_latents
+
+        padding_mask = latents.new_zeros(1, 1, height, width, dtype=transformer_dtype)
+        sigma_conditioning = torch.tensor(sigma_conditioning, dtype=torch.float32, device=device)
+        t_conditioning = self.scheduler.precondition_noise(sigma_conditioning)
+
+        # 6. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                timestep = t.view(1, 1, 1, 1, 1).expand(
+                    latents.size(0), -1, latents.size(2), -1, -1
+                )  # [B, 1, T, 1, 1]
+                current_sigma = self.scheduler.sigmas[i]
+
+                cond_latent = self.scheduler.scale_model_input(latents, t)
+                cond_latent = cond_indicator * conditioning_latents + (1 - cond_indicator) * cond_latent
+                cond_latent = cond_latent.to(transformer_dtype)
+                cond_timestep = cond_indicator * t_conditioning + (1 - cond_indicator) * timestep
+                cond_timestep = cond_timestep.to(transformer_dtype)
+
+                noise_pred = self.transformer(
+                    hidden_states=cond_latent,
+                    timestep=cond_timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    fps=fps,
+                    condition_mask=cond_mask,
+                    padding_mask=padding_mask,
+                    return_dict=False,
+                )[0]
+                noise_pred = self.scheduler.precondition_outputs(latents, noise_pred, current_sigma)
+                noise_pred = cond_indicator * conditioning_latents + (1 - cond_indicator) * noise_pred
+
+                if self.do_classifier_free_guidance:
+                    uncond_latent = self.scheduler.scale_model_input(latents, t)
+                    uncond_latent = uncond_indicator * unconditioning_latents + (1 - uncond_indicator) * uncond_latent
+                    uncond_latent = uncond_latent.to(transformer_dtype)
+                    uncond_timestep = uncond_indicator * t_conditioning + (1 - uncond_indicator) * timestep
+                    uncond_timestep = uncond_timestep.to(transformer_dtype)
+
+                    noise_pred_uncond = self.transformer(
+                        hidden_states=uncond_latent,
+                        timestep=uncond_timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        fps=fps,
+                        condition_mask=uncond_mask,
+                        padding_mask=padding_mask,
+                        return_dict=False,
+                    )[0]
+                    noise_pred_uncond = self.scheduler.precondition_outputs(latents, noise_pred_uncond, current_sigma)
+                    noise_pred_uncond = (
+                        uncond_indicator * unconditioning_latents + (1 - uncond_indicator) * noise_pred_uncond
+                    )
+                    noise_pred = noise_pred + self.guidance_scale * (noise_pred - noise_pred_uncond)
+
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, pred_original_sample=noise_pred, return_dict=False
+                )[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # 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 XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = (
+                torch.tensor(self.vae.config.latents_std)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents = latents * latents_std / self.scheduler.config.sigma_data + latents_mean
+            video = self.vae.decode(latents.to(self.vae.dtype), return_dict=False)[0]
+
+            if self.safety_checker is not None:
+                self.safety_checker.to(device)
+                video = self.video_processor.postprocess_video(video, output_type="np")
+                video = (video * 255).astype(np.uint8)
+                video_batch = []
+                for vid in video:
+                    vid = self.safety_checker.check_video_safety(vid)
+                    video_batch.append(vid)
+                video = np.stack(video_batch).astype(np.float32) / 255.0 * 2 - 1
+                video = torch.from_numpy(video).permute(0, 4, 1, 2, 3)
+                video = self.video_processor.postprocess_video(video, output_type=output_type)
+                self.safety_checker.to("cpu")
+            else:
+                video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return CosmosPipelineOutput(frames=video)

src/diffusers/pipelines/cosmos/pipeline_cosmos_text2world.py

@@ -426,12 +426,12 @@ class CosmosTextToWorldPipeline(DiffusionPipeline):
                 The height in pixels of the generated image.
             width (`int`, defaults to `1280`):
                 The width in pixels of the generated image.
-            num_frames (`int`, defaults to `129`):
+            num_frames (`int`, defaults to `121`):
                 The number of frames in the generated video.
-            num_inference_steps (`int`, defaults to `50`):
+            num_inference_steps (`int`, defaults to `36`):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
-            guidance_scale (`float`, defaults to `6.0`):
+            guidance_scale (`float`, defaults to `7.0`):
                 Guidance scale as defined in [Classifier-Free Diffusion
                 Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
                 of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
@@ -457,9 +457,6 @@ class CosmosTextToWorldPipeline(DiffusionPipeline):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
                 Whether or not to return a [`CosmosPipelineOutput`] instead of a plain tuple.
-            clip_skip (`int`, *optional*):
-                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
-                the output of the pre-final layer will be used for computing the prompt embeddings.
             callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
                 A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
                 each denoising step during the inference. with the following arguments: `callback_on_step_end(self:

src/diffusers/pipelines/cosmos/pipeline_cosmos_video2world.py

@@ -541,12 +541,12 @@ class CosmosVideoToWorldPipeline(DiffusionPipeline):
                 The height in pixels of the generated image.
             width (`int`, defaults to `1280`):
                 The width in pixels of the generated image.
-            num_frames (`int`, defaults to `129`):
+            num_frames (`int`, defaults to `121`):
                 The number of frames in the generated video.
-            num_inference_steps (`int`, defaults to `50`):
+            num_inference_steps (`int`, defaults to `36`):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
-            guidance_scale (`float`, defaults to `6.0`):
+            guidance_scale (`float`, defaults to `7.0`):
                 Guidance scale as defined in [Classifier-Free Diffusion
                 Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
                 of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
@@ -572,9 +572,6 @@ class CosmosVideoToWorldPipeline(DiffusionPipeline):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
                 Whether or not to return a [`CosmosPipelineOutput`] instead of a plain tuple.
-            clip_skip (`int`, *optional*):
-                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
-                the output of the pre-final layer will be used for computing the prompt embeddings.
             callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
                 A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
                 each denoising step during the inference. with the following arguments: `callback_on_step_end(self:

src/diffusers/pipelines/cosmos/pipeline_output.py

@@ -1,14 +1,20 @@
 from dataclasses import dataclass
+from typing import List, Union
 
+import numpy as np
+import PIL.Image
 import torch
 
-from diffusers.utils import BaseOutput
+from diffusers.utils import BaseOutput, get_logger
+
+
+logger = get_logger(__name__)
 
 
 @dataclass
 class CosmosPipelineOutput(BaseOutput):
     r"""
-    Output class for Cosmos pipelines.
+    Output class for Cosmos any-to-world/video pipelines.
 
     Args:
         frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
@@ -18,3 +24,17 @@ class CosmosPipelineOutput(BaseOutput):
     """
 
     frames: torch.Tensor
+
+
+@dataclass
+class CosmosImagePipelineOutput(BaseOutput):
+    """
+    Output class for CogView3 pipelines.
+
+    Args:
+        images (`List[PIL.Image.Image]` or `np.ndarray`)
+            List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+            num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+    """
+
+    images: Union[List[PIL.Image.Image], np.ndarray]

src/diffusers/schedulers/scheduling_cosine_dpmsolver_multistep.py

@@ -87,6 +87,7 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         lower_order_final: bool = True,
         euler_at_final: bool = False,
         final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
+        use_flow_sigmas: bool = False,
     ):
         if solver_type not in ["midpoint", "heun"]:
             if solver_type in ["logrho", "bh1", "bh2"]:
@@ -152,23 +153,19 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         if not isinstance(sigma, torch.Tensor):
             sigma = torch.tensor([sigma])
 
-        return sigma.atan() / math.pi * 2
+        if self.config.use_flow_sigmas:
+            c_noise = sigma / (sigma + 1)
+        else:
+            c_noise = sigma.atan() / math.pi * 2
+
+        return c_noise
 
     # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_outputs
     def precondition_outputs(self, sample, model_output, sigma):
-        sigma_data = self.config.sigma_data
-        c_skip = sigma_data**2 / (sigma**2 + sigma_data**2)
-
-        if self.config.prediction_type == "epsilon":
-            c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
-        elif self.config.prediction_type == "v_prediction":
-            c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+        if self.config.use_flow_sigmas:
+            return self._precondition_outputs_flow(sample, model_output, sigma)
         else:
-            raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.")
-
-        denoised = c_skip * sample + c_out * model_output
-
-        return denoised
+            return self._precondition_outputs_edm(sample, model_output, sigma)
 
     # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.scale_model_input
     def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor:
@@ -570,8 +567,42 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
 
     # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._get_conditioning_c_in
     def _get_conditioning_c_in(self, sigma):
-        c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+        if self.config.use_flow_sigmas:
+            t = sigma / (sigma + 1)
+            c_in = 1.0 - t
+        else:
+            c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
         return c_in
 
+    # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._precondition_outputs_flow
+    def _precondition_outputs_flow(self, sample, model_output, sigma):
+        t = sigma / (sigma + 1)
+        c_skip = 1.0 - t
+
+        if self.config.prediction_type == "epsilon":
+            c_out = -t
+        elif self.config.prediction_type == "v_prediction":
+            c_out = t
+        else:
+            raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.")
+
+        denoised = c_skip * sample + c_out * model_output
+        return denoised
+
+    # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._precondition_outputs_edm
+    def _precondition_outputs_edm(self, sample, model_output, sigma):
+        sigma_data = self.config.sigma_data
+        c_skip = sigma_data**2 / (sigma**2 + sigma_data**2)
+
+        if self.config.prediction_type == "epsilon":
+            c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+        elif self.config.prediction_type == "v_prediction":
+            c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+        else:
+            raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.")
+
+        denoised = c_skip * sample + c_out * model_output
+        return denoised
+
     def __len__(self):
         return self.config.num_train_timesteps

src/diffusers/schedulers/scheduling_edm_dpmsolver_multistep.py

@@ -15,14 +15,35 @@
 # DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver and https://github.com/NVlabs/edm
 
 import math
+from dataclasses import dataclass
 from typing import List, Optional, Tuple, Union
 
 import numpy as np
 import torch
 
 from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
 from ..utils.torch_utils import randn_tensor
-from .scheduling_utils import SchedulerMixin, SchedulerOutput
+from .scheduling_utils import SchedulerMixin
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EDMDPMSolverMultistep
+class EDMDPMSolverMultistepSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
+            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+            denoising loop.
+        pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
+            The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+            `pred_original_sample` can be used to preview progress or for guidance.
+    """
+
+    prev_sample: torch.Tensor
+    pred_original_sample: Optional[torch.Tensor] = None
 
 
 class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
@@ -107,6 +128,7 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         lower_order_final: bool = True,
         euler_at_final: bool = False,
         final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
+        use_flow_sigmas: bool = False,
     ):
         # settings for DPM-Solver
         if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"]:
@@ -185,25 +207,19 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         if not isinstance(sigma, torch.Tensor):
             sigma = torch.tensor([sigma])
 
-        c_noise = 0.25 * torch.log(sigma)
+        if self.config.use_flow_sigmas:
+            c_noise = sigma / (sigma + 1)
+        else:
+            c_noise = 0.25 * torch.log(sigma)
 
         return c_noise
 
     # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_outputs
     def precondition_outputs(self, sample, model_output, sigma):
-        sigma_data = self.config.sigma_data
-        c_skip = sigma_data**2 / (sigma**2 + sigma_data**2)
-
-        if self.config.prediction_type == "epsilon":
-            c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
-        elif self.config.prediction_type == "v_prediction":
-            c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+        if self.config.use_flow_sigmas:
+            return self._precondition_outputs_flow(sample, model_output, sigma)
         else:
-            raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.")
-
-        denoised = c_skip * sample + c_out * model_output
-
-        return denoised
+            return self._precondition_outputs_edm(sample, model_output, sigma)
 
     # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.scale_model_input
     def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor:
@@ -598,7 +614,8 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         sample: torch.Tensor,
         generator=None,
         return_dict: bool = True,
-    ) -> Union[SchedulerOutput, Tuple]:
+        pred_original_sample: Optional[torch.Tensor] = None,
+    ) -> Union[EDMDPMSolverMultistepSchedulerOutput, Tuple]:
         """
         Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
         the multistep DPMSolver.
@@ -613,12 +630,14 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
             generator (`torch.Generator`, *optional*):
                 A random number generator.
             return_dict (`bool`):
-                Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
+                Whether or not to return a
+                [`~schedulers.scheduling_edm_dpmsolver_multistep.EDMDPMSolverMultistepSchedulerOutput`] or a `tuple`.
 
         Returns:
-            [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
-                If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
-                tuple is returned where the first element is the sample tensor.
+            [`~schedulers.scheduling_edm_dpmsolver_multistep.EDMDPMSolverMultistepSchedulerOutput`] or `tuple`:
+                If return_dict is `True`,
+                [`~schedulers.scheduling_edm_dpmsolver_multistep.EDMDPMSolverMultistepSchedulerOutput`] is returned,
+                otherwise a tuple is returned where the first element is the sample tensor.
 
         """
         if self.num_inference_steps is None:
@@ -639,7 +658,12 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
             (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15
         )
 
-        model_output = self.convert_model_output(model_output, sample=sample)
+        if pred_original_sample is None:
+            model_output = self.convert_model_output(model_output, sample=sample)
+        else:
+            model_output = pred_original_sample
+            # TODO: thresholding is not handled in this case, but probably not needed either for Cosmos
+
         for i in range(self.config.solver_order - 1):
             self.model_outputs[i] = self.model_outputs[i + 1]
         self.model_outputs[-1] = model_output
@@ -667,7 +691,7 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         if not return_dict:
             return (prev_sample,)
 
-        return SchedulerOutput(prev_sample=prev_sample)
+        return EDMDPMSolverMultistepSchedulerOutput(prev_sample=prev_sample, pred_original_sample=model_output)
 
     # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise
     def add_noise(
@@ -705,8 +729,42 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
 
     # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._get_conditioning_c_in
     def _get_conditioning_c_in(self, sigma):
-        c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+        if self.config.use_flow_sigmas:
+            t = sigma / (sigma + 1)
+            c_in = 1.0 - t
+        else:
+            c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
         return c_in
 
+    # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._precondition_outputs_flow
+    def _precondition_outputs_flow(self, sample, model_output, sigma):
+        t = sigma / (sigma + 1)
+        c_skip = 1.0 - t
+
+        if self.config.prediction_type == "epsilon":
+            c_out = -t
+        elif self.config.prediction_type == "v_prediction":
+            c_out = t
+        else:
+            raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.")
+
+        denoised = c_skip * sample + c_out * model_output
+        return denoised
+
+    # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._precondition_outputs_edm
+    def _precondition_outputs_edm(self, sample, model_output, sigma):
+        sigma_data = self.config.sigma_data
+        c_skip = sigma_data**2 / (sigma**2 + sigma_data**2)
+
+        if self.config.prediction_type == "epsilon":
+            c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+        elif self.config.prediction_type == "v_prediction":
+            c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+        else:
+            raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.")
+
+        denoised = c_skip * sample + c_out * model_output
+        return denoised
+
     def __len__(self):
         return self.config.num_train_timesteps

src/diffusers/schedulers/scheduling_edm_euler.py

@@ -28,7 +28,7 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
 @dataclass
-# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EulerDiscrete
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EDMEuler
 class EDMEulerSchedulerOutput(BaseOutput):
     """
     Output class for the scheduler's `step` function output.
@@ -96,6 +96,7 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
         prediction_type: str = "epsilon",
         rho: float = 7.0,
         final_sigmas_type: str = "zero",  # can be "zero" or "sigma_min"
+        use_flow_sigmas: bool = False,
     ):
         if sigma_schedule not in ["karras", "exponential"]:
             raise ValueError(f"Wrong value for provided for `{sigma_schedule=}`.`")
@@ -169,24 +170,18 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
         if not isinstance(sigma, torch.Tensor):
             sigma = torch.tensor([sigma])
 
-        c_noise = 0.25 * torch.log(sigma)
+        if self.config.use_flow_sigmas:
+            c_noise = sigma / (sigma + 1)
+        else:
+            c_noise = 0.25 * torch.log(sigma)
 
         return c_noise
 
     def precondition_outputs(self, sample, model_output, sigma):
-        sigma_data = self.config.sigma_data
-        c_skip = sigma_data**2 / (sigma**2 + sigma_data**2)
-
-        if self.config.prediction_type == "epsilon":
-            c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
-        elif self.config.prediction_type == "v_prediction":
-            c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+        if self.config.use_flow_sigmas:
+            return self._precondition_outputs_flow(sample, model_output, sigma)
         else:
-            raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.")
-
-        denoised = c_skip * sample + c_out * model_output
-
-        return denoised
+            return self._precondition_outputs_edm(sample, model_output, sigma)
 
     def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor:
         """
@@ -441,8 +436,40 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
         return noisy_samples
 
     def _get_conditioning_c_in(self, sigma):
-        c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+        if self.config.use_flow_sigmas:
+            t = sigma / (sigma + 1)
+            c_in = 1.0 - t
+        else:
+            c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
         return c_in
 
+    def _precondition_outputs_flow(self, sample, model_output, sigma):
+        t = sigma / (sigma + 1)
+        c_skip = 1.0 - t
+
+        if self.config.prediction_type == "epsilon":
+            c_out = -t
+        elif self.config.prediction_type == "v_prediction":
+            c_out = t
+        else:
+            raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.")
+
+        denoised = c_skip * sample + c_out * model_output
+        return denoised
+
+    def _precondition_outputs_edm(self, sample, model_output, sigma):
+        sigma_data = self.config.sigma_data
+        c_skip = sigma_data**2 / (sigma**2 + sigma_data**2)
+
+        if self.config.prediction_type == "epsilon":
+            c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+        elif self.config.prediction_type == "v_prediction":
+            c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+        else:
+            raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.")
+
+        denoised = c_skip * sample + c_out * model_output
+        return denoised
+
     def __len__(self):
         return self.config.num_train_timesteps

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -407,6 +407,36 @@ class ConsisIDPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class Cosmos2TextToImagePipeline(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 Cosmos2VideoToWorldPipeline(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 CosmosTextToWorldPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/cosmos/test_cosmos2_text2_image.py

@@ -0,0 +1,342 @@
+# Copyright 2024 The HuggingFace 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.
+
+import inspect
+import json
+import os
+import tempfile
+import unittest
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, T5EncoderModel
+
+from diffusers import AutoencoderKLWan, Cosmos2TextToImagePipeline, CosmosTransformer3DModel, EDMEulerScheduler
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin, to_np
+from .cosmos_guardrail import DummyCosmosSafetyChecker
+
+
+enable_full_determinism()
+
+
+class Cosmos2TextToImagePipelineWrapper(Cosmos2TextToImagePipeline):
+    @staticmethod
+    def from_pretrained(*args, **kwargs):
+        kwargs["safety_checker"] = DummyCosmosSafetyChecker()
+        return Cosmos2TextToImagePipeline.from_pretrained(*args, **kwargs)
+
+
+class Cosmos2TextToImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = Cosmos2TextToImagePipelineWrapper
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    supports_dduf = False
+    test_xformers_attention = False
+    test_layerwise_casting = True
+    test_group_offloading = True
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = CosmosTransformer3DModel(
+            in_channels=16,
+            out_channels=16,
+            num_attention_heads=2,
+            attention_head_dim=16,
+            num_layers=2,
+            mlp_ratio=2,
+            text_embed_dim=32,
+            adaln_lora_dim=4,
+            max_size=(4, 32, 32),
+            patch_size=(1, 2, 2),
+            rope_scale=(2.0, 1.0, 1.0),
+            concat_padding_mask=True,
+            extra_pos_embed_type="learnable",
+        )
+
+        torch.manual_seed(0)
+        vae = AutoencoderKLWan(
+            base_dim=3,
+            z_dim=16,
+            dim_mult=[1, 1, 1, 1],
+            num_res_blocks=1,
+            temperal_downsample=[False, True, True],
+        )
+
+        torch.manual_seed(0)
+        scheduler = EDMEulerScheduler(
+            sigma_min=0.002,
+            sigma_max=80,
+            sigma_data=0.5,
+            sigma_schedule="karras",
+            num_train_timesteps=1000,
+            prediction_type="epsilon",
+            rho=7.0,
+            final_sigmas_type="sigma_min",
+            use_flow_sigmas=True,
+        )
+        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            # We cannot run the Cosmos Guardrail for fast tests due to the large model size
+            "safety_checker": DummyCosmosSafetyChecker(),
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        inputs = {
+            "prompt": "dance monkey",
+            "negative_prompt": "bad quality",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 3.0,
+            "height": 32,
+            "width": 32,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+        }
+
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        generated_image = image[0]
+
+        self.assertEqual(generated_image.shape, (3, 32, 32))
+        expected_video = torch.randn(3, 32, 32)
+        max_diff = np.abs(generated_image - expected_video).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-2)
+
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
+            )
+
+    def test_vae_tiling(self, expected_diff_max: float = 0.2):
+        generator_device = "cpu"
+        components = self.get_dummy_components()
+
+        pipe = self.pipeline_class(**components)
+        pipe.to("cpu")
+        pipe.set_progress_bar_config(disable=None)
+
+        # Without tiling
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_without_tiling = pipe(**inputs)[0]
+
+        # With tiling
+        pipe.vae.enable_tiling(
+            tile_sample_min_height=96,
+            tile_sample_min_width=96,
+            tile_sample_stride_height=64,
+            tile_sample_stride_width=64,
+        )
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_with_tiling = pipe(**inputs)[0]
+
+        self.assertLess(
+            (to_np(output_without_tiling) - to_np(output_with_tiling)).max(),
+            expected_diff_max,
+            "VAE tiling should not affect the inference results",
+        )
+
+    def test_save_load_optional_components(self, expected_max_difference=1e-4):
+        self.pipeline_class._optional_components.remove("safety_checker")
+        super().test_save_load_optional_components(expected_max_difference=expected_max_difference)
+        self.pipeline_class._optional_components.append("safety_checker")
+
+    def test_serialization_with_variants(self):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        model_components = [
+            component_name
+            for component_name, component in pipe.components.items()
+            if isinstance(component, torch.nn.Module)
+        ]
+        model_components.remove("safety_checker")
+        variant = "fp16"
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            pipe.save_pretrained(tmpdir, variant=variant, safe_serialization=False)
+
+            with open(f"{tmpdir}/model_index.json", "r") as f:
+                config = json.load(f)
+
+            for subfolder in os.listdir(tmpdir):
+                if not os.path.isfile(subfolder) and subfolder in model_components:
+                    folder_path = os.path.join(tmpdir, subfolder)
+                    is_folder = os.path.isdir(folder_path) and subfolder in config
+                    assert is_folder and any(p.split(".")[1].startswith(variant) for p in os.listdir(folder_path))
+
+    def test_torch_dtype_dict(self):
+        components = self.get_dummy_components()
+        if not components:
+            self.skipTest("No dummy components defined.")
+
+        pipe = self.pipeline_class(**components)
+
+        specified_key = next(iter(components.keys()))
+
+        with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
+            pipe.save_pretrained(tmpdirname, safe_serialization=False)
+            torch_dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
+            loaded_pipe = self.pipeline_class.from_pretrained(
+                tmpdirname, safety_checker=DummyCosmosSafetyChecker(), torch_dtype=torch_dtype_dict
+            )
+
+        for name, component in loaded_pipe.components.items():
+            if name == "safety_checker":
+                continue
+            if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
+                expected_dtype = torch_dtype_dict.get(name, torch_dtype_dict.get("default", torch.float32))
+                self.assertEqual(
+                    component.dtype,
+                    expected_dtype,
+                    f"Component '{name}' has dtype {component.dtype} but expected {expected_dtype}",
+                )
+
+    @unittest.skip(
+        "The pipeline should not be runnable without a safety checker. The test creates a pipeline without passing in "
+        "a safety checker, which makes the pipeline default to the actual Cosmos Guardrail. The Cosmos Guardrail is "
+        "too large and slow to run on CI."
+    )
+    def test_encode_prompt_works_in_isolation(self):
+        pass

tests/pipelines/cosmos/test_cosmos2_video2world.py

@@ -0,0 +1,356 @@
+# Copyright 2024 The HuggingFace 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.
+
+import inspect
+import json
+import os
+import tempfile
+import unittest
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import AutoTokenizer, T5EncoderModel
+
+from diffusers import AutoencoderKLWan, Cosmos2VideoToWorldPipeline, CosmosTransformer3DModel, EDMEulerScheduler
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin, to_np
+from .cosmos_guardrail import DummyCosmosSafetyChecker
+
+
+enable_full_determinism()
+
+
+class Cosmos2VideoToWorldPipelineWrapper(Cosmos2VideoToWorldPipeline):
+    @staticmethod
+    def from_pretrained(*args, **kwargs):
+        kwargs["safety_checker"] = DummyCosmosSafetyChecker()
+        return Cosmos2VideoToWorldPipeline.from_pretrained(*args, **kwargs)
+
+
+class Cosmos2VideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = Cosmos2VideoToWorldPipelineWrapper
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS.union({"image", "video"})
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    supports_dduf = False
+    test_xformers_attention = False
+    test_layerwise_casting = True
+    test_group_offloading = True
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = CosmosTransformer3DModel(
+            in_channels=16 + 1,
+            out_channels=16,
+            num_attention_heads=2,
+            attention_head_dim=16,
+            num_layers=2,
+            mlp_ratio=2,
+            text_embed_dim=32,
+            adaln_lora_dim=4,
+            max_size=(4, 32, 32),
+            patch_size=(1, 2, 2),
+            rope_scale=(2.0, 1.0, 1.0),
+            concat_padding_mask=True,
+            extra_pos_embed_type="learnable",
+        )
+
+        torch.manual_seed(0)
+        vae = AutoencoderKLWan(
+            base_dim=3,
+            z_dim=16,
+            dim_mult=[1, 1, 1, 1],
+            num_res_blocks=1,
+            temperal_downsample=[False, True, True],
+        )
+
+        torch.manual_seed(0)
+        scheduler = EDMEulerScheduler(
+            sigma_min=0.002,
+            sigma_max=80,
+            sigma_data=0.5,
+            sigma_schedule="karras",
+            num_train_timesteps=1000,
+            prediction_type="epsilon",
+            rho=7.0,
+            final_sigmas_type="sigma_min",
+            use_flow_sigmas=True,
+        )
+        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            # We cannot run the Cosmos Guardrail for fast tests due to the large model size
+            "safety_checker": DummyCosmosSafetyChecker(),
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        image_height = 32
+        image_width = 32
+        image = PIL.Image.new("RGB", (image_width, image_height))
+
+        inputs = {
+            "image": image,
+            "prompt": "dance monkey",
+            "negative_prompt": "bad quality",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 3.0,
+            "height": image_height,
+            "width": image_width,
+            "num_frames": 9,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+        }
+
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+
+        self.assertEqual(generated_video.shape, (9, 3, 32, 32))
+        expected_video = torch.randn(9, 3, 32, 32)
+        max_diff = np.abs(generated_video - expected_video).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_components_function(self):
+        init_components = self.get_dummy_components()
+        init_components = {k: v for k, v in init_components.items() if not isinstance(v, (str, int, float))}
+        pipe = self.pipeline_class(**init_components)
+        self.assertTrue(hasattr(pipe, "components"))
+        self.assertTrue(set(pipe.components.keys()) == set(init_components.keys()))
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-2)
+
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
+            )
+
+    def test_vae_tiling(self, expected_diff_max: float = 0.2):
+        generator_device = "cpu"
+        components = self.get_dummy_components()
+
+        pipe = self.pipeline_class(**components)
+        pipe.to("cpu")
+        pipe.set_progress_bar_config(disable=None)
+
+        # Without tiling
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_without_tiling = pipe(**inputs)[0]
+
+        # With tiling
+        pipe.vae.enable_tiling(
+            tile_sample_min_height=96,
+            tile_sample_min_width=96,
+            tile_sample_stride_height=64,
+            tile_sample_stride_width=64,
+        )
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_with_tiling = pipe(**inputs)[0]
+
+        self.assertLess(
+            (to_np(output_without_tiling) - to_np(output_with_tiling)).max(),
+            expected_diff_max,
+            "VAE tiling should not affect the inference results",
+        )
+
+    def test_save_load_optional_components(self, expected_max_difference=1e-4):
+        self.pipeline_class._optional_components.remove("safety_checker")
+        super().test_save_load_optional_components(expected_max_difference=expected_max_difference)
+        self.pipeline_class._optional_components.append("safety_checker")
+
+    def test_serialization_with_variants(self):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        model_components = [
+            component_name
+            for component_name, component in pipe.components.items()
+            if isinstance(component, torch.nn.Module)
+        ]
+        model_components.remove("safety_checker")
+        variant = "fp16"
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            pipe.save_pretrained(tmpdir, variant=variant, safe_serialization=False)
+
+            with open(f"{tmpdir}/model_index.json", "r") as f:
+                config = json.load(f)
+
+            for subfolder in os.listdir(tmpdir):
+                if not os.path.isfile(subfolder) and subfolder in model_components:
+                    folder_path = os.path.join(tmpdir, subfolder)
+                    is_folder = os.path.isdir(folder_path) and subfolder in config
+                    assert is_folder and any(p.split(".")[1].startswith(variant) for p in os.listdir(folder_path))
+
+    def test_torch_dtype_dict(self):
+        components = self.get_dummy_components()
+        if not components:
+            self.skipTest("No dummy components defined.")
+
+        pipe = self.pipeline_class(**components)
+
+        specified_key = next(iter(components.keys()))
+
+        with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
+            pipe.save_pretrained(tmpdirname, safe_serialization=False)
+            torch_dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
+            loaded_pipe = self.pipeline_class.from_pretrained(
+                tmpdirname, safety_checker=DummyCosmosSafetyChecker(), torch_dtype=torch_dtype_dict
+            )
+
+        for name, component in loaded_pipe.components.items():
+            if name == "safety_checker":
+                continue
+            if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
+                expected_dtype = torch_dtype_dict.get(name, torch_dtype_dict.get("default", torch.float32))
+                self.assertEqual(
+                    component.dtype,
+                    expected_dtype,
+                    f"Component '{name}' has dtype {component.dtype} but expected {expected_dtype}",
+                )
+
+    @unittest.skip(
+        "The pipeline should not be runnable without a safety checker. The test creates a pipeline without passing in "
+        "a safety checker, which makes the pipeline default to the actual Cosmos Guardrail. The Cosmos Guardrail is "
+        "too large and slow to run on CI."
+    )
+    def test_encode_prompt_works_in_isolation(self):
+        pass