[Ernie-Image] Add lora support

[asomoza]

What does this PR do?

Adds initial lora support for Ernie-Image:

import torch
from diffusers import ErnieImagePipeline

pipe = ErnieImagePipeline.from_pretrained("baidu/ERNIE-Image-Turbo", torch_dtype=torch.bfloat16)
pipe.load_lora_weights("HexWave/Ernie_Lora_pack", weight_name="Yumikon.safetensors")
pipe.enable_model_cpu_offload()

prompt = "yum1kon, a capybara wearing a jacket and riding a skate in the city street."
images = pipe(
    prompt=prompt,
    height=1264,
    width=848,
    num_inference_steps=8,
    guidance_scale=1.0,
    generator=torch.Generator("cuda").manual_seed(42),
    use_pe=False,
).images
images[0].save("ernie-image-turbo-output.png")

w/o lora	lora

Fixes #13501

Who can review?

@sayakpaul @SamuelTallet @kayselmecnun

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[asomoza]

@claude do a review

[github-actions]

Claude finished @asomoza's task in 3m 19s —— View job

---

Review of PR #13575: [Ernie-Image] Add lora support

• Read review guidelines (AGENTS.md, models.md, pipelines.md)
• Read the PR diff
• Review src/diffusers/loaders/__init__.py
• Review src/diffusers/loaders/lora_pipeline.py
• Review src/diffusers/models/transformers/transformer_ernie_image.py
• Review src/diffusers/pipelines/ernie_image/pipeline_ernie_image.py
• Check top-level exports
• Post review feedback

---

Overall

Clean, well-structured PR. The implementation correctly follows the established LoRA loader mixin pattern (copies from CogVideoXLoraLoaderMixin / SD3LoraLoaderMixin), the model and pipeline inheritance changes are correct, and the # Copied from annotations are properly set. The state dict key transformation logic is appropriately customized for Ernie-Image's architecture. No major issues found.

Detailed Findings

1. src/diffusers/loaders/lora_pipeline.py — ErnieImageLoraLoaderMixin ✅

• lora_state_dict: The PEFT → diffusion_model.* → transformer.* key remapping chain is correct and the comment at lines 5896–5898 explaining why only the prefix needs to change (module paths already match diffusers naming) is a nice touch.
• load_lora_weights (line 5906): Correctly # Copied from CogVideoXLoraLoaderMixin.load_lora_weights.
• load_lora_into_transformer (line 5949): Correctly # Copied from SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->ErnieImageTransformer2DModel.
• save_lora_weights, fuse_lora, unfuse_lora: All correctly # Copied from their CogVideoX counterparts.

2. src/diffusers/models/transformers/transformer_ernie_image.py ✅

• Adding PeftAdapterMixin to the MRO of ErnieImageTransformer2DModel (line 292) is the correct pattern for enabling LoRA adapter loading on the transformer.
• MRO order (ModelMixin, ConfigMixin, PeftAdapterMixin) is consistent with other transformer models.

3. src/diffusers/pipelines/ernie_image/pipeline_ernie_image.py ✅

• Adding ErnieImageLoraLoaderMixin to the pipeline's MRO (line 35) is correct.
• MRO order (DiffusionPipeline, ErnieImageLoraLoaderMixin) matches the established pattern.

4. src/diffusers/loaders/__init__.py ✅

• Properly registered in both _import_structure["lora_pipeline"] (line 88) and the TYPE_CHECKING block (line 121), in alphabetical order.

5. docs/source/en/api/loaders/lora.md ✅

• Correctly placed. The [[autodoc]] section is in the right spot.

Nits (non-blocking)

None for this PR, but I noticed a pre-existing typo on line 38 of docs/source/en/api/loaders/lora.md — LTX2LoraLoaderMixin says "provides similar functions for [Flux2]" instead of "LTX2". Not from this PR but worth a follow-up fix.

Verdict

LGTM — the implementation is clean, follows conventions, and I don't see any correctness issues. 👍

---

[sayakpaul]

So good!