This repository provides a native implementation of the PuLID (Perceptual Lighting Image Diffusion) model for ComfyUI, a popular node-based interface for Stable Diffusion. It aims to offer users a way to control image lighting and style based on a reference image directly within the ComfyUI ecosystem, potentially improving workflow efficiency for generative AI artists.

How It Works

The implementation leverages a 4-step lighting UNet, mirroring the original PuLID approach. It integrates with ComfyUI's node system, allowing users to connect PuLID functionality into their existing workflows. The project emphasizes the importance of reference image quality and suggests experimenting with CFG, samplers (like sgm_uniform), and the method parameter for controlling fidelity versus style. An "Advanced" node offers a fidelity slider and projection options for finer control.

Quick Start & Requirements

• Installation: Place PuLID pre-trained models in ComfyUI/models/pulid/. Install facexlib and InsightFace with AntelopeV2 models (place unzipped models in ComfyUI/models/insightface/models/antelopev2). EVA CLIP models are downloaded automatically.
• Prerequisites: ComfyUI, facexlib, InsightFace (AntelopeV2). IPAdapter extension must be updated for compatibility.
• Resources: Requires downloading pre-trained models.

Highlighted Details

• Faithful implementation compared to the official Gradio demo.
• Offers method parameter for fidelity/style control, including a neutral option.
• "Advanced" node provides a fidelity slider and projection option.
• IPAdapter compatibility requires an extension update.

Maintenance & Community

The repository is in "maintenance only" mode as of April 14, 2025. The author may merge crucial updates or PRs but does not plan consistent work.

Licensing & Compatibility

The README does not explicitly state a license. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The code is considered beta, with potential for changes. The "Lightning lora" does not work as well, and quality degradation may occur with certain models. The effectiveness is highly dependent on the quality and sharpness of the reference image.