ComfyUI LongCat-Image Integration

This custom node integrates the LongCat-Image pipeline into ComfyUI, enabling text-to-image generation and image editing with the LongCat-Image models.

Features

• Text-to-Image Generation: Generate high-quality images from text prompts using LongCat-Image models
• Image Editing: Edit existing images with instruction-based prompts using LongCat-Image-Edit models
• Chinese Text Support: Excellent Chinese text rendering capabilities
• Efficient: Only 6B parameters with competitive performance

Installation

1. Install Dependencies

cd custom_nodes/comfyui_longcat_image
pip install -r requirements.txt

2. Install LongCat-Image Package

pip install git+https://github.com/meituan-longcat/LongCat-Image.git

3. Download Models

Download the models using huggingface-cli:

pip install "huggingface_hub[cli]"

# For text-to-image
huggingface-cli download meituan-longcat/LongCat-Image --local-dir models/diffusion_models/LongCat-Image

# For image editing
huggingface-cli download meituan-longcat/LongCat-Image-Edit --local-dir models/diffusion_models/LongCat-Image-Edit

# For fine-tuning (optional)
huggingface-cli download meituan-longcat/LongCat-Image-Dev --local-dir models/diffusion_models/LongCat-Image-Dev

Available Nodes

LongCat-Image Model Loader

Loads a LongCat-Image model for use with other nodes.

Inputs:

• model_path: Path to the model directory (e.g., "LongCat-Image" or "LongCat-Image-Edit")
• dtype: Data type for model weights (bfloat16, float16, float32)

Outputs:

• LONGCAT_PIPE: Pipeline object for use with generation nodes

LongCat-Image Text to Image

Generates images from text prompts.

Inputs:

• LONGCAT_PIPE: Pipeline from the model loader
• prompt: Text description of the image to generate
• negative_prompt: Things to avoid in the generated image
• width: Image width (default: 1344)
• height: Image height (default: 768)
• steps: Number of inference steps (default: 50)
• guidance_scale: CFG scale (default: 4.5)
• seed: Random seed
• enable_cfg_renorm: Enable CFG renormalization (true/false)
• enable_prompt_rewrite: Enable built-in prompt rewriting (true/false)

Outputs:

• IMAGE: Generated image

LongCat-Image Edit

Edits images based on instruction prompts.

Inputs:

• LONGCAT_PIPE: Pipeline from the model loader (must be an edit model)
• image: Input image to edit
• prompt: Edit instruction
• negative_prompt: Things to avoid in the edited image
• steps: Number of inference steps (default: 50)
• guidance_scale: CFG scale (default: 4.5)
• seed: Random seed

Outputs:

• IMAGE: Edited image

Example Workflows

Example workflow JSON files are provided in this directory:

• example_workflow_t2i.json - Text-to-image generation workflow
• example_workflow_edit.json - Image editing workflow

You can load these workflows in ComfyUI by dragging and dropping the JSON file onto the canvas.

Text-to-Image

1. Add a LongCat-Image Model Loader node

   • Set model_path to "LongCat-Image"

2. Add a LongCat-Image Text to Image node

   • Connect the loader output to the pipeline input
   • Enter your prompt
   • Adjust settings as needed

3. Add a Save Image node to save the output

Image Editing

1. Add a LongCat-Image Model Loader node

   • Set model_path to "LongCat-Image-Edit"

2. Add a Load Image node to load your input image

3. Add a LongCat-Image Edit node

   • Connect the loader output to the pipeline input
   • Connect the image to edit
   • Enter your edit instruction (e.g., "将猫变成狗" - "change the cat to a dog")

4. Add a Save Image node to save the output

Model Information

Model	Type	Description
LongCat-Image	Text-to-Image	Final release model for out-of-the-box inference
LongCat-Image-Dev	Text-to-Image	Mid-training checkpoint, suitable for fine-tuning
LongCat-Image-Edit	Image Editing	Specialized model for image editing

Performance

• Parameters: 6B (highly efficient)
• Supported Resolutions: 768x1344 and variations
• Chinese Text Support: Industry-leading Chinese dictionary coverage
• Quality: Competitive with much larger models

Tips

• For better results, use a strong LLM for prompt engineering
• The model has excellent Chinese text rendering capabilities
• Enable prompt rewriting for enhanced generation quality
• Default guidance scale of 4.5 works well for most cases

License

LongCat-Image is licensed under Apache 2.0. See the LongCat-Image repository for more information.

References

• LongCat-Image GitHub
• LongCat-Image on Hugging Face
• Technical Report