Add Chroma modular pipeline (text2image)

Contribution to huggingface#13295. Adds a ModularPipeline for Chroma (t2i only):
- T5-only text encoder block with Chroma's padding-token attention masking
- guider-based CFG denoise loop (ClassifierFreeGuidance)
- reuses Flux-style packed-latent prep, mu-shift timesteps, RoPE ids, decode

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dcg23vUZRMfkX5mkfdKh3Q

Author: claude

docs/source/en/api/pipelines/chroma.md

@@ -105,3 +105,11 @@ image.save("chroma-single-file.png")
 [[autodoc]] ChromaInpaintPipeline
   - all
   - __call__
+
+## ChromaModularPipeline
+
+[[autodoc]] ChromaModularPipeline
+
+## ChromaAutoBlocks
+
+[[autodoc]] ChromaAutoBlocks

src/diffusers/__init__.py

@@ -503,6 +503,8 @@
         [
             "AnimaAutoBlocks",
             "AnimaModularPipeline",
+            "ChromaAutoBlocks",
+            "ChromaModularPipeline",
             "Cosmos3DistilledBlocks",
             "Cosmos3DistilledModularPipeline",
             "Cosmos3OmniBlocks",
@@ -1329,6 +1331,8 @@
         from .modular_pipelines import (
             AnimaAutoBlocks,
             AnimaModularPipeline,
+            ChromaAutoBlocks,
+            ChromaModularPipeline,
             Cosmos3DistilledBlocks,
             Cosmos3DistilledModularPipeline,
             Cosmos3OmniBlocks,

src/diffusers/loaders/peft.py

@@ -46,7 +46,7 @@
 logger = logging.get_logger(__name__)
 
 _SET_ADAPTER_SCALE_FN_MAPPING = defaultdict(
-    lambda: (lambda model_cls, weights: weights),
+    lambda: lambda model_cls, weights: weights,
     {
         "UNet2DConditionModel": _maybe_expand_lora_scales,
         "UNetMotionModel": _maybe_expand_lora_scales,

src/diffusers/modular_pipelines/__init__.py

@@ -103,6 +103,10 @@
         "AnimaAutoBlocks",
         "AnimaModularPipeline",
     ]
+    _import_structure["chroma"] = [
+        "ChromaAutoBlocks",
+        "ChromaModularPipeline",
+    ]
     _import_structure["cosmos"] = [
         "Cosmos3DistilledBlocks",
         "Cosmos3DistilledModularPipeline",
@@ -139,6 +143,7 @@
         from ..utils.dummy_pt_objects import *  # noqa F403
     else:
         from .anima import AnimaAutoBlocks, AnimaModularPipeline
+        from .chroma import ChromaAutoBlocks, ChromaModularPipeline
         from .components_manager import ComponentsManager
         from .cosmos import (
             Cosmos3DistilledBlocks,

src/diffusers/modular_pipelines/anima/modular_blocks_anima.py

@@ -34,8 +34,10 @@ class AnimaCoreDenoiseStep(SequentialPipelineBlocks):
     Denoise block that takes encoded Anima text inputs and runs the denoising process.
 
       Components:
-          text_conditioner (`AnimaTextConditioner`) transformer (`CosmosTransformer3DModel`) scheduler
-          (`FlowMatchEulerDiscreteScheduler`) guider (`ClassifierFreeGuidance`)
+          text_conditioner (`AnimaTextConditioner`)
+          transformer (`CosmosTransformer3DModel`)
+          scheduler (`FlowMatchEulerDiscreteScheduler`)
+          guider (`ClassifierFreeGuidance`)
 
       Inputs:
           qwen_prompt_embeds (`Tensor`):
@@ -100,7 +102,8 @@ class AnimaDecodeStep(SequentialPipelineBlocks):
     Decode Anima latents into generated images.
 
       Components:
-          vae (`AutoencoderKLQwenImage`) image_processor (`VaeImageProcessor`)
+          vae (`AutoencoderKLQwenImage`)
+          image_processor (`VaeImageProcessor`)
 
       Inputs:
           latents (`Tensor`):
@@ -128,12 +131,13 @@ def outputs(self):
 # auto_docstring
 class AnimaImg2ImgCoreDenoiseStep(SequentialPipelineBlocks):
     """
-    Denoise block for Anima image-to-image generation. Uses image_latents already in state from
-    AnimaImg2ImgVaeEncoderStep.
+    Denoise block for Anima image-to-image generation. Uses image_latents already in state from AnimaImg2ImgVaeEncoderStep.
 
       Components:
-          text_conditioner (`AnimaTextConditioner`) transformer (`CosmosTransformer3DModel`) scheduler
-          (`FlowMatchEulerDiscreteScheduler`) guider (`ClassifierFreeGuidance`)
+          text_conditioner (`AnimaTextConditioner`)
+          transformer (`CosmosTransformer3DModel`)
+          scheduler (`FlowMatchEulerDiscreteScheduler`)
+          guider (`ClassifierFreeGuidance`)
 
       Inputs:
           qwen_prompt_embeds (`Tensor`):
@@ -203,13 +207,13 @@ def outputs(self):
 # auto_docstring
 class AnimaAutoCoreDenoiseStep(AutoPipelineBlocks):
     """
-    Denoise step that selects between text-to-image and image-to-image denoising based on whether image_latents is
-    present in state. - `AnimaCoreDenoiseStep` (text2image) is used when no image_latents are present. -
-    `AnimaImg2ImgCoreDenoiseStep` (img2img) is used when image_latents are present.
+    Denoise step that selects between text-to-image and image-to-image denoising based on whether image_latents is present in state. - `AnimaCoreDenoiseStep` (text2image) is used when no image_latents are present. - `AnimaImg2ImgCoreDenoiseStep` (img2img) is used when image_latents are present.
 
       Components:
-          text_conditioner (`AnimaTextConditioner`) transformer (`CosmosTransformer3DModel`) scheduler
-          (`FlowMatchEulerDiscreteScheduler`) guider (`ClassifierFreeGuidance`)
+          text_conditioner (`AnimaTextConditioner`)
+          transformer (`CosmosTransformer3DModel`)
+          scheduler (`FlowMatchEulerDiscreteScheduler`)
+          guider (`ClassifierFreeGuidance`)
 
       Inputs:
           qwen_prompt_embeds (`Tensor`):
@@ -271,11 +275,11 @@ def description(self) -> str:
 # auto_docstring
 class AnimaAutoVaeImageEncoderStep(AutoPipelineBlocks):
     """
-    VAE Image Encoder step that encodes the input image to produce image_latents. Skipped when no image is provided
-    (text-to-image workflow).
+    VAE Image Encoder step that encodes the input image to produce image_latents. Skipped when no image is provided (text-to-image workflow).
 
       Components:
-          vae (`AutoencoderKLQwenImage`) image_processor (`VaeImageProcessor`)
+          vae (`AutoencoderKLQwenImage`)
+          image_processor (`VaeImageProcessor`)
 
       Inputs:
           image (`Image | list`, *optional*):
@@ -318,10 +322,15 @@ class AnimaAutoBlocks(SequentialPipelineBlocks):
         - `img2img`: requires `image`, `prompt`
 
       Components:
-          text_encoder (`Qwen3Model`) tokenizer (`Qwen2Tokenizer`) t5_tokenizer (`T5Tokenizer`) guider
-          (`ClassifierFreeGuidance`) vae (`AutoencoderKLQwenImage`) image_processor (`VaeImageProcessor`)
-          text_conditioner (`AnimaTextConditioner`) transformer (`CosmosTransformer3DModel`) scheduler
-          (`FlowMatchEulerDiscreteScheduler`)
+          text_encoder (`Qwen3Model`)
+          tokenizer (`Qwen2Tokenizer`)
+          t5_tokenizer (`T5Tokenizer`)
+          guider (`ClassifierFreeGuidance`)
+          vae (`AutoencoderKLQwenImage`)
+          image_processor (`VaeImageProcessor`)
+          text_conditioner (`AnimaTextConditioner`)
+          transformer (`CosmosTransformer3DModel`)
+          scheduler (`FlowMatchEulerDiscreteScheduler`)
 
       Inputs:
           prompt (`str`):

src/diffusers/modular_pipelines/chroma/__init__.py

@@ -0,0 +1,47 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["modular_blocks_chroma"] = ["ChromaAutoBlocks"]
+    _import_structure["modular_pipeline"] = ["ChromaModularPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *  # noqa F403
+    else:
+        from .modular_blocks_chroma import ChromaAutoBlocks
+        from .modular_pipeline import ChromaModularPipeline
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)