support use_flow_sigmas in EDM scheduler instead of maintain cosmos-specific scheduler

Author: a-r-r-o-w

scripts/convert_cosmos_to_diffusers.py

@@ -14,7 +14,6 @@
     CosmosTextToWorldPipeline,
     CosmosTransformer3DModel,
     EDMEulerScheduler,
-    FlowMatchEulerEDMCosmos2_0Scheduler,
 )
 
 
@@ -187,6 +186,51 @@ def rename_transformer_blocks_(key: str, state_dict: Dict[str, Any]):
         "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": (20 / 24, 2.0, 2.0),
+        "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, 1.0, 1.0),
+        "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 = {
@@ -352,8 +396,8 @@ def convert_vae(vae_type: str):
     return vae
 
 
-def save_pipeline_cosmos_1_0(args, transformer, vae, dtype):
-    text_encoder = T5EncoderModel.from_pretrained(args.text_encoder_path, torch_dtype=dtype)
+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.
@@ -378,11 +422,11 @@ def save_pipeline_cosmos_1_0(args, transformer, vae, dtype):
     pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")
 
 
-def save_pipeline_cosmos_2_0(args, transformer, vae, dtype):
-    text_encoder = T5EncoderModel.from_pretrained(args.text_encoder_path, torch_dtype=dtype)
+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 = FlowMatchEulerEDMCosmos2_0Scheduler(
+    scheduler = EDMEulerScheduler(
         sigma_min=0.0002,
         sigma_max=80,
         sigma_data=1.0,
@@ -391,6 +435,7 @@ def save_pipeline_cosmos_2_0(args, transformer, vae, dtype):
         prediction_type="epsilon",
         rho=7.0,
         final_sigmas_type="sigma_min",
+        use_flow_sigmas=True,
     )
 
     pipe = CosmosTextToImagePipeline(
@@ -458,8 +503,8 @@ def get_args():
 
     if args.save_pipeline:
         if "Cosmos-1.0" in args.transformer_type:
-            save_pipeline_cosmos_1_0(args, transformer, vae, dtype)
+            save_pipeline_cosmos_1_0(args, transformer, vae)
         elif "Cosmos-2.0" in args.transformer_type:
-            save_pipeline_cosmos_2_0(args, transformer, vae, dtype)
+            save_pipeline_cosmos_2_0(args, transformer, vae)
         else:
             assert False

src/diffusers/__init__.py

@@ -271,7 +271,6 @@
             "EulerAncestralDiscreteScheduler",
             "EulerDiscreteScheduler",
             "FlowMatchEulerDiscreteScheduler",
-            "FlowMatchEulerEDMCosmos2_0Scheduler",
             "FlowMatchHeunDiscreteScheduler",
             "FlowMatchLCMScheduler",
             "HeunDiscreteScheduler",
@@ -880,7 +879,6 @@
             EulerAncestralDiscreteScheduler,
             EulerDiscreteScheduler,
             FlowMatchEulerDiscreteScheduler,
-            FlowMatchEulerEDMCosmos2_0Scheduler,
             FlowMatchHeunDiscreteScheduler,
             FlowMatchLCMScheduler,
             HeunDiscreteScheduler,

src/diffusers/pipelines/cosmos/pipeline_cosmos_text2image.py

@@ -21,7 +21,7 @@
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...models import AutoencoderKLWan, CosmosTransformer3DModel
-from ...schedulers import FlowMatchEulerEDMCosmos2_0Scheduler
+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
@@ -56,15 +56,17 @@ def __init__(self, *args, **kwargs):
         >>> import torch
         >>> from diffusers import CosmosTextToImagePipeline
 
-        >>> # TODO(aryan): update model_id
-        >>> model_id = "/raid/aryan/diffusers-cosmos2-t2i-2B"
+        >>> # Available checkpoints: nvidia/Cosmos-Predict2-2B-Text2Image, nvidia/Cosmos-Predict2-14B-Text2Image
+        >>> model_id = "nvidia/Cosmos-Predict2-2B-Text2Image"
         >>> pipe = CosmosTextToImagePipeline.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, height=1024, width=1024, generator=torch.Generator().manual_seed(1)).images[0]
+        >>> output = pipe(
+        ...     prompt=prompt, negative_prompt=negative_prompt, generator=torch.Generator().manual_seed(1)
+        ... ).images[0]
         >>> output.save("output.png")
         ```
 """
@@ -147,7 +149,7 @@ class CosmosTextToImagePipeline(DiffusionPipeline):
             [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
         transformer ([`CosmosTransformer3DModel`]):
             Conditional Transformer to denoise the encoded image latents.
-        scheduler ([`FlowMatchEulerEDMCosmos2_0Scheduler`]):
+        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.
@@ -164,7 +166,7 @@ def __init__(
         tokenizer: T5TokenizerFast,
         transformer: CosmosTransformer3DModel,
         vae: AutoencoderKLWan,
-        scheduler: FlowMatchEulerEDMCosmos2_0Scheduler,
+        scheduler: EDMEulerScheduler,
         safety_checker: CosmosSafetyChecker = None,
     ):
         super().__init__()
@@ -228,13 +230,13 @@ def _get_t5_prompt_embeds(
 
         return prompt_embeds
 
-    # Copied from diffusers.pipelines.cosmos.pipeline_cosmos_text2world.CosmosTextToWorldPipeline.encode_prompt
+    # 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_videos_per_prompt: int = 1,
+        num_images_per_prompt: int = 1,
         prompt_embeds: Optional[torch.Tensor] = None,
         negative_prompt_embeds: Optional[torch.Tensor] = None,
         max_sequence_length: int = 512,
@@ -253,7 +255,7 @@ def encode_prompt(
                 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):
+            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
@@ -282,8 +284,8 @@ def encode_prompt(
 
             # 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)
+            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 ""
@@ -307,8 +309,8 @@ def encode_prompt(
 
             # 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)
+            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
 
@@ -402,7 +404,7 @@ def __call__(
         width: int = 1280,
         num_inference_steps: int = 35,
         guidance_scale: float = 7.0,
-        num_videos_per_prompt: Optional[int] = 1,
+        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,
@@ -434,7 +436,7 @@ def __call__(
                 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_videos_per_prompt (`int`, *optional*, defaults to 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
@@ -521,7 +523,7 @@ def __call__(
             prompt=prompt,
             negative_prompt=negative_prompt,
             do_classifier_free_guidance=self.do_classifier_free_guidance,
-            num_videos_per_prompt=num_videos_per_prompt,
+            num_images_per_prompt=num_images_per_prompt,
             prompt_embeds=prompt_embeds,
             negative_prompt_embeds=negative_prompt_embeds,
             device=device,
@@ -535,7 +537,7 @@ def __call__(
         transformer_dtype = self.transformer.dtype
         num_channels_latents = self.transformer.config.in_channels
         latents = self.prepare_latents(
-            batch_size * num_videos_per_prompt,
+            batch_size * num_images_per_prompt,
             num_channels_latents,
             height,
             width,
@@ -612,7 +614,7 @@ def __call__(
             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 + latents_mean
+            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:

src/diffusers/schedulers/__init__.py

@@ -59,7 +59,6 @@
     _import_structure["scheduling_euler_ancestral_discrete"] = ["EulerAncestralDiscreteScheduler"]
     _import_structure["scheduling_euler_discrete"] = ["EulerDiscreteScheduler"]
     _import_structure["scheduling_flow_match_euler_discrete"] = ["FlowMatchEulerDiscreteScheduler"]
-    _import_structure["scheduling_flow_match_euler_edm_cosmos2"] = ["FlowMatchEulerEDMCosmos2_0Scheduler"]
     _import_structure["scheduling_flow_match_heun_discrete"] = ["FlowMatchHeunDiscreteScheduler"]
     _import_structure["scheduling_flow_match_lcm"] = ["FlowMatchLCMScheduler"]
     _import_structure["scheduling_heun_discrete"] = ["HeunDiscreteScheduler"]
@@ -162,7 +161,6 @@
         from .scheduling_euler_ancestral_discrete import EulerAncestralDiscreteScheduler
         from .scheduling_euler_discrete import EulerDiscreteScheduler
         from .scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
-        from .scheduling_flow_match_euler_edm_cosmos2 import FlowMatchEulerEDMCosmos2_0Scheduler
         from .scheduling_flow_match_heun_discrete import FlowMatchHeunDiscreteScheduler
         from .scheduling_flow_match_lcm import FlowMatchLCMScheduler
         from .scheduling_heun_discrete import HeunDiscreteScheduler

src/diffusers/schedulers/scheduling_edm_euler.py

@@ -96,6 +96,7 @@ def __init__(
         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 @@ def precondition_noise(self, sigma):
         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 @@ def add_noise(
         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