ComfyUI-CLIPSkip

A custom node for ComfyUI that adds CLIP skip functionality to vanilla WAN workflow using CLIP. This node allows you to skip a specified number of layers in a CLIP model, which can adjust the style or quality of image embeddings in generation pipelines.

Installation

Via Git (Recommended)

1. Open a terminal in your ComfyUI custom_nodes directory:

cd ComfyUI/custom_nodes

2. Clone this repository:

git clone https://github.com/yourusername/ComfyUI-CLIPSkip.git

3. Restart ComfyUI. The node will be automatically loaded.

4. Ensure you have the required WAN CLIP model (e.g., umt5_xxl_fp8_e4m3fn_scaled.safetensors) in ComfyUI/models/text_encoders/.

Manual Installation

1. Download this repository as a ZIP file.
2. Extract it into the ComfyUI/custom_nodes directory.
3. Rename the folder to ComfyUI-CLIPSkip if needed.
4. Restart ComfyUI.

Dependencies

• ComfyUI (latest version recommended)
• PyTorch (installed with ComfyUI)

No additional dependencies are required.

Usage

1. Load a CLIP Vision model using CLIPVisionLoader or any other node that outputs CLIP_VISION.
2. Connect the clip_vision output to the clip input of CLIPSkip.
3. Set the skip_layers parameter (e.g., 1 to skip the last layer, 0 to disable skipping).
4. Connect the output clip to any node that accepts CLIP_VISION (e.g., CLIPVisionEncode).

Example Workflow

CLIPVisionLoader -> CLIPSkip -> CLIPVisionEncode -> (further pipeline)

Supported Models

• umt5_xxl_fp8_e4m3fn_scaled.safetensors (24 layers).

Notes

• Designed for WAN-type CLIP models in ComfyUI.
• Requires ComfyUI with FP8 support for optimal performance.

License

MIT License (see LICENSE file for details).

Contributing

Feel free to submit issues or pull requests on GitHub!