ComfyUI_RH_ZImageI2L

ComfyUI Z-Image I2L (Image to LoRA)

ComfyUI custom nodes for Z-Image Image-to-LoRA generation. Generate personalized LoRA weights from reference images using DiffSynth-Studio's Z-Image pipeline.

✨ Features

• Image to LoRA: Generate LoRA weights directly from reference images
• No Training Required: Instant LoRA generation without traditional fine-tuning
• ComfyUI Integration: Seamless workflow integration with standard LoRA nodes

🛠️ Installation

1. Clone this repository into your ComfyUI custom nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/HM-RunningHub/ComfyUI_RH_ZImageI2L.git

2. Install dependencies:

pip install -r requirements.txt

📦 Model Downloads

Models will be automatically downloaded from ModelScope on first run and cached locally.

Required Models

| Model | Description | Files | |-------|-------------|-------| | Tongyi-MAI/Z-Image | Base transformer | transformer/*.safetensors | | Tongyi-MAI/Z-Image-Turbo | Text encoder, VAE & Tokenizer | text_encoder/*.safetensors, vae/, tokenizer/ | | DiffSynth-Studio/General-Image-Encoders | Image encoders | SigLIP2-G384/, DINOv3-7B/ | | DiffSynth-Studio/Z-Image-i2L | Image to LoRA model | model.safetensors |

Model Cache Path

Models are automatically downloaded and cached in the ModelScope cache directory:

| OS | Default Cache Path | |----|-------------------| | Linux | ~/.cache/modelscope/hub/ | | Windows | C:\Users\<username>\.cache\modelscope\hub\ | | macOS | ~/.cache/modelscope/hub/ |

The directory structure after download:

~/.cache/modelscope/hub/
├── Tongyi-MAI/
│   ├── Z-Image/
│   │   └── transformer/*.safetensors
│   └── Z-Image-Turbo/
│       ├── text_encoder/*.safetensors
│       ├── vae/diffusion_pytorch_model.safetensors
│       └── tokenizer/
└── DiffSynth-Studio/
    ├── General-Image-Encoders/
    │   ├── SigLIP2-G384/model.safetensors
    │   └── DINOv3-7B/model.safetensors
    └── Z-Image-i2L/
        └── model.safetensors

Custom Cache Directory

You can customize the cache directory by setting the MODELSCOPE_CACHE environment variable:

# Linux/macOS
export MODELSCOPE_CACHE=/path/to/your/cache

# Windows (PowerShell)
$env:MODELSCOPE_CACHE = "D:\models\modelscope"

# Windows (CMD)
set MODELSCOPE_CACHE=D:\models\modelscope

🚀 Usage

Nodes

| Node | Description | |------|-------------| | ZImageI2L Loader | Load the Z-Image I2L pipeline | | ZImageI2L LoRA Generator | Generate LoRA from input images | | ZImageI2L Saver | Save generated LoRA to output folder |

Basic Workflow

1. Add ZImageI2L Loader to load the pipeline
2. Connect your reference images to ZImageI2L LoRA Generator
3. Use ZImageI2L Saver to save the generated LoRA
4. Use the generated LoRA with any standard LoRA loader node

Example Workflow

An example API workflow is provided in the workflows folder:

• workflows/zimage_i2l_example_api.json - Complete workflow demonstrating LoRA generation and usage

This workflow includes:

• Loading reference images (4 images)
• Generating LoRA with ZImageI2L nodes
• Applying the generated LoRA to Z-Image model
• Generating images with the personalized LoRA

📝 Parameters

ZImageI2L LoRA Generator

| Parameter | Type | Description | |-----------|------|-------------| | pipeline | RH_ZImageI2LPipeline | Pipeline from Loader node | | training_images | IMAGE | Reference images for LoRA generation | | seed | INT | Random seed for reproducibility |

⚠️ Requirements

• VRAM: 24GB+ recommended (tested on RTX 4090)
• Python: 3.10+
• ComfyUI: Latest version

🙏 Acknowledgments

• DiffSynth-Studio - Core Z-Image pipeline
• Tongyi-MAI - Z-Image models

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.