ComfyUI RH QwenImage I2L (Image-to-LoRA)

A ComfyUI custom node that generates an Image-to-LoRA (I2L) LoRA from one or more training images using DiffSynth-Studio Qwen-Image i2L pipelines.

• Nodes
  • RunningHub_ImageQwenI2L_Loader(Style)
  • RunningHub_ImageQwenI2L_Loader(CFB)
  • RunningHub_ImageQwenI2L_LoraGenerator (outputs a .safetensors LoRA into ComfyUI/models/loras/)

Original model page: DiffSynth-Studio/Qwen-Image-i2L on ModelScope

---

Installation

1) Install this ComfyUI node

Clone into your ComfyUI custom_nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/HM-RunningHub/ComfyUI_RH_QwenImageI2L.git

Install Python dependencies:

cd ComfyUI/custom_nodes/ComfyUI_RH_QwenImageI2L
pip install -r requirements.txt

Note: sync.sh is intentionally ignored and should not be committed.

2) Install DiffSynth-Studio (required)

This project uses diffsynth from DiffSynth-Studio. Install the latest version:

git clone https://github.com/modelscope/DiffSynth-Studio.git
cd DiffSynth-Studio
pip install -e .

3) (Recommended) Install git-lfs

Some model repositories require Git LFS:

git lfs install

4) Restart ComfyUI

After installation, restart ComfyUI to load the new nodes.

---

Models (Required)

This node loads models via DiffSynth-Studio model configs. On first run, models may be auto-downloaded (requires network access to the model host used by DiffSynth/ModelScope/HF).

If you prefer offline / deterministic setup, place models under your ComfyUI models directory.

Recommended offline layout

Put the following under ComfyUI/models/:

ComfyUI/models/
├─ DiffSynth-Studio/
│  ├─ General-Image-Encoders/
│  │  ├─ SigLIP2-G384/model.safetensors
│  │  └─ DINOv3-7B/model.safetensors
│  └─ Qwen-Image-i2L/
│     ├─ Qwen-Image-i2L-Style.safetensors
│     ├─ Qwen-Image-i2L-Coarse.safetensors
│     ├─ Qwen-Image-i2L-Fine.safetensors
│     └─ Qwen-Image-i2L-Bias.safetensors
└─ Qwen/
   ├─ Qwen-Image/
   │  └─ text_encoder/model*.safetensors
   └─ Qwen-Image-Edit/
      └─ processor/

What each Loader needs

• Loader(Style) loads:

  • DiffSynth-Studio/General-Image-Encoders
    • SigLIP2-G384/model.safetensors
    • DINOv3-7B/model.safetensors
  • DiffSynth-Studio/Qwen-Image-i2L
    • Qwen-Image-i2L-Style.safetensors
  • Processor:
    • Qwen/Qwen-Image-Edit → processor/

• Loader(CFB) loads:

  • Qwen/Qwen-Image
    • text_encoder/model*.safetensors
  • DiffSynth-Studio/General-Image-Encoders
    • SigLIP2-G384/model.safetensors
    • DINOv3-7B/model.safetensors
  • DiffSynth-Studio/Qwen-Image-i2L
    • Qwen-Image-i2L-Coarse.safetensors
    • Qwen-Image-i2L-Fine.safetensors
    • (optional at generation time) Qwen-Image-i2L-Bias.safetensors
  • Processor:
    • Qwen/Qwen-Image-Edit → processor/

---

Usage (ComfyUI)

Basic workflow (Style)

[Load Image(s)] → [RunningHub_ImageQwenI2L_Loader(Style)] → [RunningHub_ImageQwenI2L_LoraGenerator] → (LoRA saved to models/loras)

Output

RunningHub_ImageQwenI2L_LoraGenerator writes a LoRA file:

• Path: ComfyUI/models/loras/i2l_style_lora_<uuid>.safetensors
• Returns:
  • lora_name (for ComfyUI LoRA dropdown)
  • lora_path (full path)

---

Notes

• GPU: the pipeline uses cuda and bfloat16. Make sure your GPU/driver supports BF16.
• Disk offload: the loader uses disk offload to reduce VRAM usage; ensure you have enough free disk space.

---

License

Open-source. If you redistribute, please comply with the licenses of upstream projects and model providers.