Chroma Pipeline

docs/source/en/_toctree.yml

@@ -283,6 +283,8 @@
         title: AllegroTransformer3DModel
       - local: api/models/aura_flow_transformer2d
         title: AuraFlowTransformer2DModel
+      - local: api/models/chroma_transformer
+        title: ChromaTransformer2DModel
       - local: api/models/cogvideox_transformer3d
         title: CogVideoXTransformer3DModel
       - local: api/models/cogview3plus_transformer2d
@@ -405,6 +407,8 @@
       title: AutoPipeline
     - local: api/pipelines/blip_diffusion
       title: BLIP-Diffusion
+    - local: api/pipelines/chroma
+      title: Chroma
     - local: api/pipelines/cogvideox
       title: CogVideoX
     - local: api/pipelines/cogview3

docs/source/en/api/models/chroma_transformer.md

@@ -0,0 +1,19 @@
+<!--Copyright 2025 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.
+-->
+
+# ChromaTransformer2DModel
+
+A modified flux Transformer model from [Chroma](https://huggingface.co/lodestones/Chroma)
+
+## ChromaTransformer2DModel
+
+[[autodoc]] ChromaTransformer2DModel

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

@@ -0,0 +1,71 @@
+<!--Copyright 2025 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.
+-->
+
+# Chroma
+
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
+</div>
+
+Chroma is a text to image generation model based on Flux.
+
+Original model checkpoints for Chroma can be found [here](https://huggingface.co/lodestones/Chroma).
+
+<Tip>
+
+Chroma can use all the same optimizations as Flux.
+
+</Tip>
+
+## Inference (Single File)
+
+The `ChromaTransformer2DModel` supports loading checkpoints in the original format. This is also useful when trying to load finetunes or quantized versions of the models that have been published by the community.
+
+The following example demonstrates how to run Chroma from a single file.
+
+Then run the following example
+
+```python
+import torch
+from diffusers import ChromaTransformer2DModel, ChromaPipeline
+from transformers import T5EncoderModel
+
+bfl_repo = "black-forest-labs/FLUX.1-dev"
+dtype = torch.bfloat16
+
+transformer = ChromaTransformer2DModel.from_single_file("https://huggingface.co/lodestones/Chroma/blob/main/chroma-unlocked-v35.safetensors", torch_dtype=dtype)
+
+text_encoder = T5EncoderModel.from_pretrained(bfl_repo, subfolder="text_encoder_2", torch_dtype=dtype)
+tokenizer = T5Tokenizer.from_pretrained(bfl_repo, subfolder="tokenizer_2", torch_dtype=dtype)
+
+pipe = ChromaPipeline.from_pretrained(bfl_repo, transformer=transformer, text_encoder=text_encoder, tokenizer=tokenizer, torch_dtype=dtype)
+
+pipe.enable_model_cpu_offload()
+
+prompt = "A cat holding a sign that says hello world"
+image = pipe(
+    prompt,
+    guidance_scale=4.0,
+    output_type="pil",
+    num_inference_steps=26,
+    generator=torch.Generator("cpu").manual_seed(0)
+).images[0]
+
+image.save("image.png")
+```
+
+## ChromaPipeline
+
+[[autodoc]] ChromaPipeline
+	- all
+	- __call__

src/diffusers/__init__.py

@@ -159,6 +159,7 @@
             "AutoencoderTiny",
             "AutoModel",
             "CacheMixin",
+            "ChromaTransformer2DModel",
             "CogVideoXTransformer3DModel",
             "CogView3PlusTransformer2DModel",
             "CogView4Transformer2DModel",
@@ -352,6 +353,7 @@
             "AuraFlowPipeline",
             "BlipDiffusionControlNetPipeline",
             "BlipDiffusionPipeline",
+            "ChromaPipeline",
             "CLIPImageProjection",
             "CogVideoXFunControlPipeline",
             "CogVideoXImageToVideoPipeline",
@@ -768,6 +770,7 @@
             AutoencoderTiny,
             AutoModel,
             CacheMixin,
+            ChromaTransformer2DModel,
             CogVideoXTransformer3DModel,
             CogView3PlusTransformer2DModel,
             CogView4Transformer2DModel,
@@ -940,6 +943,7 @@
             AudioLDM2UNet2DConditionModel,
             AudioLDMPipeline,
             AuraFlowPipeline,
+            ChromaPipeline,
             CLIPImageProjection,
             CogVideoXFunControlPipeline,
             CogVideoXImageToVideoPipeline,

src/diffusers/loaders/peft.py

@@ -60,6 +60,7 @@
     "HiDreamImageTransformer2DModel": lambda model_cls, weights: weights,
     "HunyuanVideoFramepackTransformer3DModel": lambda model_cls, weights: weights,
     "WanVACETransformer3DModel": lambda model_cls, weights: weights,
+    "ChromaTransformer2DModel": lambda model_cls, weights: weights,
 }
 
 

src/diffusers/loaders/single_file_model.py

@@ -29,6 +29,7 @@
     convert_animatediff_checkpoint_to_diffusers,
     convert_auraflow_transformer_checkpoint_to_diffusers,
     convert_autoencoder_dc_checkpoint_to_diffusers,
+    convert_chroma_transformer_checkpoint_to_diffusers,
     convert_controlnet_checkpoint,
     convert_flux_transformer_checkpoint_to_diffusers,
     convert_hidream_transformer_to_diffusers,
@@ -97,6 +98,10 @@
         "checkpoint_mapping_fn": convert_flux_transformer_checkpoint_to_diffusers,
         "default_subfolder": "transformer",
     },
+    "ChromaTransformer2DModel": {
+        "checkpoint_mapping_fn": convert_chroma_transformer_checkpoint_to_diffusers,
+        "default_subfolder": "transformer",
+    },
     "LTXVideoTransformer3DModel": {
         "checkpoint_mapping_fn": convert_ltx_transformer_checkpoint_to_diffusers,
         "default_subfolder": "transformer",

src/diffusers/loaders/single_file_utils.py

@@ -3310,3 +3310,172 @@ def convert_hidream_transformer_to_diffusers(checkpoint, **kwargs):
             checkpoint[k.replace("model.diffusion_model.", "")] = checkpoint.pop(k)
 
     return checkpoint
+
+
+def convert_chroma_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
+    converted_state_dict = {}
+    keys = list(checkpoint.keys())
+
+    for k in keys:
+        if "model.diffusion_model." in k:
+            checkpoint[k.replace("model.diffusion_model.", "")] = checkpoint.pop(k)
+
+    num_layers = list(set(int(k.split(".", 2)[1]) for k in checkpoint if "double_blocks." in k))[-1] + 1  # noqa: C401
+    num_single_layers = list(set(int(k.split(".", 2)[1]) for k in checkpoint if "single_blocks." in k))[-1] + 1  # noqa: C401
+    num_guidance_layers = (
+        list(set(int(k.split(".", 3)[2]) for k in checkpoint if "distilled_guidance_layer.layers." in k))[-1] + 1  # noqa: C401
+    )
+    mlp_ratio = 4.0
+    inner_dim = 3072
+
+    # in SD3 original implementation of AdaLayerNormContinuous, it split linear projection output into shift, scale;
+    # while in diffusers it split into scale, shift. Here we swap the linear projection weights in order to be able to use diffusers implementation
+    def swap_scale_shift(weight):
+        shift, scale = weight.chunk(2, dim=0)
+        new_weight = torch.cat([scale, shift], dim=0)
+        return new_weight
+
+    # guidance
+    converted_state_dict["distilled_guidance_layer.in_proj.bias"] = checkpoint.pop(
+        "distilled_guidance_layer.in_proj.bias"
+    )
+    converted_state_dict["distilled_guidance_layer.in_proj.weight"] = checkpoint.pop(
+        "distilled_guidance_layer.in_proj.weight"
+    )
+    converted_state_dict["distilled_guidance_layer.out_proj.bias"] = checkpoint.pop(
+        "distilled_guidance_layer.out_proj.bias"
+    )
+    converted_state_dict["distilled_guidance_layer.out_proj.weight"] = checkpoint.pop(
+        "distilled_guidance_layer.out_proj.weight"
+    )
+    for i in range(num_guidance_layers):
+        block_prefix = f"distilled_guidance_layer.layers.{i}."
+        converted_state_dict[f"{block_prefix}linear_1.bias"] = checkpoint.pop(
+            f"distilled_guidance_layer.layers.{i}.in_layer.bias"
+        )
+        converted_state_dict[f"{block_prefix}linear_1.weight"] = checkpoint.pop(
+            f"distilled_guidance_layer.layers.{i}.in_layer.weight"
+        )
+        converted_state_dict[f"{block_prefix}linear_2.bias"] = checkpoint.pop(
+            f"distilled_guidance_layer.layers.{i}.out_layer.bias"
+        )
+        converted_state_dict[f"{block_prefix}linear_2.weight"] = checkpoint.pop(
+            f"distilled_guidance_layer.layers.{i}.out_layer.weight"
+        )
+        converted_state_dict[f"distilled_guidance_layer.norms.{i}.weight"] = checkpoint.pop(
+            f"distilled_guidance_layer.norms.{i}.scale"
+        )
+
+    # context_embedder
+    converted_state_dict["context_embedder.weight"] = checkpoint.pop("txt_in.weight")
+    converted_state_dict["context_embedder.bias"] = checkpoint.pop("txt_in.bias")
+
+    # x_embedder
+    converted_state_dict["x_embedder.weight"] = checkpoint.pop("img_in.weight")
+    converted_state_dict["x_embedder.bias"] = checkpoint.pop("img_in.bias")
+
+    # double transformer blocks
+    for i in range(num_layers):
+        block_prefix = f"transformer_blocks.{i}."
+        # Q, K, V
+        sample_q, sample_k, sample_v = torch.chunk(checkpoint.pop(f"double_blocks.{i}.img_attn.qkv.weight"), 3, dim=0)
+        context_q, context_k, context_v = torch.chunk(
+            checkpoint.pop(f"double_blocks.{i}.txt_attn.qkv.weight"), 3, dim=0
+        )
+        sample_q_bias, sample_k_bias, sample_v_bias = torch.chunk(
+            checkpoint.pop(f"double_blocks.{i}.img_attn.qkv.bias"), 3, dim=0
+        )
+        context_q_bias, context_k_bias, context_v_bias = torch.chunk(
+            checkpoint.pop(f"double_blocks.{i}.txt_attn.qkv.bias"), 3, dim=0
+        )
+        converted_state_dict[f"{block_prefix}attn.to_q.weight"] = torch.cat([sample_q])
+        converted_state_dict[f"{block_prefix}attn.to_q.bias"] = torch.cat([sample_q_bias])
+        converted_state_dict[f"{block_prefix}attn.to_k.weight"] = torch.cat([sample_k])
+        converted_state_dict[f"{block_prefix}attn.to_k.bias"] = torch.cat([sample_k_bias])
+        converted_state_dict[f"{block_prefix}attn.to_v.weight"] = torch.cat([sample_v])
+        converted_state_dict[f"{block_prefix}attn.to_v.bias"] = torch.cat([sample_v_bias])
+        converted_state_dict[f"{block_prefix}attn.add_q_proj.weight"] = torch.cat([context_q])
+        converted_state_dict[f"{block_prefix}attn.add_q_proj.bias"] = torch.cat([context_q_bias])
+        converted_state_dict[f"{block_prefix}attn.add_k_proj.weight"] = torch.cat([context_k])
+        converted_state_dict[f"{block_prefix}attn.add_k_proj.bias"] = torch.cat([context_k_bias])
+        converted_state_dict[f"{block_prefix}attn.add_v_proj.weight"] = torch.cat([context_v])
+        converted_state_dict[f"{block_prefix}attn.add_v_proj.bias"] = torch.cat([context_v_bias])
+        # qk_norm
+        converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.img_attn.norm.query_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.img_attn.norm.key_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_added_q.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_attn.norm.query_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_added_k.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_attn.norm.key_norm.scale"
+        )
+        # ff img_mlp
+        converted_state_dict[f"{block_prefix}ff.net.0.proj.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.img_mlp.0.weight"
+        )
+        converted_state_dict[f"{block_prefix}ff.net.0.proj.bias"] = checkpoint.pop(f"double_blocks.{i}.img_mlp.0.bias")
+        converted_state_dict[f"{block_prefix}ff.net.2.weight"] = checkpoint.pop(f"double_blocks.{i}.img_mlp.2.weight")
+        converted_state_dict[f"{block_prefix}ff.net.2.bias"] = checkpoint.pop(f"double_blocks.{i}.img_mlp.2.bias")
+        converted_state_dict[f"{block_prefix}ff_context.net.0.proj.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_mlp.0.weight"
+        )
+        converted_state_dict[f"{block_prefix}ff_context.net.0.proj.bias"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_mlp.0.bias"
+        )
+        converted_state_dict[f"{block_prefix}ff_context.net.2.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_mlp.2.weight"
+        )
+        converted_state_dict[f"{block_prefix}ff_context.net.2.bias"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_mlp.2.bias"
+        )
+        # output projections.
+        converted_state_dict[f"{block_prefix}attn.to_out.0.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.img_attn.proj.weight"
+        )
+        converted_state_dict[f"{block_prefix}attn.to_out.0.bias"] = checkpoint.pop(
+            f"double_blocks.{i}.img_attn.proj.bias"
+        )
+        converted_state_dict[f"{block_prefix}attn.to_add_out.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_attn.proj.weight"
+        )
+        converted_state_dict[f"{block_prefix}attn.to_add_out.bias"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_attn.proj.bias"
+        )
+
+    # single transformer blocks
+    for i in range(num_single_layers):
+        block_prefix = f"single_transformer_blocks.{i}."
+        # Q, K, V, mlp
+        mlp_hidden_dim = int(inner_dim * mlp_ratio)
+        split_size = (inner_dim, inner_dim, inner_dim, mlp_hidden_dim)
+        q, k, v, mlp = torch.split(checkpoint.pop(f"single_blocks.{i}.linear1.weight"), split_size, dim=0)
+        q_bias, k_bias, v_bias, mlp_bias = torch.split(
+            checkpoint.pop(f"single_blocks.{i}.linear1.bias"), split_size, dim=0
+        )
+        converted_state_dict[f"{block_prefix}attn.to_q.weight"] = torch.cat([q])
+        converted_state_dict[f"{block_prefix}attn.to_q.bias"] = torch.cat([q_bias])
+        converted_state_dict[f"{block_prefix}attn.to_k.weight"] = torch.cat([k])
+        converted_state_dict[f"{block_prefix}attn.to_k.bias"] = torch.cat([k_bias])
+        converted_state_dict[f"{block_prefix}attn.to_v.weight"] = torch.cat([v])
+        converted_state_dict[f"{block_prefix}attn.to_v.bias"] = torch.cat([v_bias])
+        converted_state_dict[f"{block_prefix}proj_mlp.weight"] = torch.cat([mlp])
+        converted_state_dict[f"{block_prefix}proj_mlp.bias"] = torch.cat([mlp_bias])
+        # qk norm
+        converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = checkpoint.pop(
+            f"single_blocks.{i}.norm.query_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = checkpoint.pop(
+            f"single_blocks.{i}.norm.key_norm.scale"
+        )
+        # output projections.
+        converted_state_dict[f"{block_prefix}proj_out.weight"] = checkpoint.pop(f"single_blocks.{i}.linear2.weight")
+        converted_state_dict[f"{block_prefix}proj_out.bias"] = checkpoint.pop(f"single_blocks.{i}.linear2.bias")
+
+    converted_state_dict["proj_out.weight"] = checkpoint.pop("final_layer.linear.weight")
+    converted_state_dict["proj_out.bias"] = checkpoint.pop("final_layer.linear.bias")
+
+    return converted_state_dict

src/diffusers/models/__init__.py

@@ -74,6 +74,7 @@
     _import_structure["transformers.t5_film_transformer"] = ["T5FilmDecoder"]
     _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
     _import_structure["transformers.transformer_allegro"] = ["AllegroTransformer3DModel"]
+    _import_structure["transformers.transformer_chroma"] = ["ChromaTransformer2DModel"]
     _import_structure["transformers.transformer_cogview3plus"] = ["CogView3PlusTransformer2DModel"]
     _import_structure["transformers.transformer_cogview4"] = ["CogView4Transformer2DModel"]
     _import_structure["transformers.transformer_cosmos"] = ["CosmosTransformer3DModel"]
@@ -151,6 +152,7 @@
         from .transformers import (
             AllegroTransformer3DModel,
             AuraFlowTransformer2DModel,
+            ChromaTransformer2DModel,
             CogVideoXTransformer3DModel,
             CogView3PlusTransformer2DModel,
             CogView4Transformer2DModel,

src/diffusers/models/embeddings.py

@@ -31,7 +31,7 @@ def get_timestep_embedding(
     downscale_freq_shift: float = 1,
     scale: float = 1,
     max_period: int = 10000,
-):
+) -> torch.Tensor:
     """
     This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings.
 
@@ -1325,7 +1325,7 @@ def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shif
         self.downscale_freq_shift = downscale_freq_shift
         self.scale = scale
 
-    def forward(self, timesteps):
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
         t_emb = get_timestep_embedding(
             timesteps,
             self.num_channels,

src/diffusers/models/transformers/__init__.py

@@ -17,6 +17,7 @@
     from .t5_film_transformer import T5FilmDecoder
     from .transformer_2d import Transformer2DModel
     from .transformer_allegro import AllegroTransformer3DModel
+    from .transformer_chroma import ChromaTransformer2DModel
     from .transformer_cogview3plus import CogView3PlusTransformer2DModel
     from .transformer_cogview4 import CogView4Transformer2DModel
     from .transformer_cosmos import CosmosTransformer3DModel