This ComfyUI custom node implements PuLID-Flux, a tuning-free ID customization solution for FLUX.1-dev models. It targets users of ComfyUI looking to integrate advanced image generation and editing capabilities, offering a prototype for testing model quality with convenient model hacks.

How It Works

The implementation leverages model hacks for the encoder section of Flux.1-dev models. It supports various model formats including 32bit/16bit, 8bit GGUF, and FP8 (e5m2, e4m3fn), catering to different VRAM requirements. The node aims to provide a more formal reimplementation of PuLID for Flux, with current development focusing on weight, start_at, and end_at parameters.

Quick Start & Requirements

• Install by cloning the repository into ComfyUI/custom_nodes.
• Install dependencies from requirements.txt.
• Requires Flux.1-dev models (unet, clip, encoder) placed in respective ComfyUI model directories.
• EVA CLIP (EVA02-CLIP-L-14-336) is auto-downloaded or requires manual placement in ComfyUI/models/clip.
• Requires facexlib and InsightFace with AntelopeV2 models placed in ComfyUI/models/insightface/models/antelopev2.
• GGUF models require ComfyUI-GGUF.
• Recommended for models with CUDA compute >= v8.0 for FP8 Flux models.

Highlighted Details

• Supports weight, start_at, and end_at parameters.
• Offers compatibility with multiple Flux.1-dev model quantizations (32bit, 16bit, 8bit GGUF, FP8).
• Inspired by cubiq/PuLID_ComfyUI.

Maintenance & Community

• Alpha version, actively under development.
• Credits ComfyUI, PuLID for Flux, and cubiq/PuLID_ComfyUI.

Licensing & Compatibility

• License not explicitly stated in the README.
• Compatibility with commercial or closed-source projects is not specified.

Limitations & Caveats

The ApplyPulidFlux node does not work on hardware with CUDA compute less than v8.0. The attn_mask functionality is not yet implemented. Disconnecting the ApplyPulidFlux node after its first run does not fully reset the Flux model's influence.