One-image style cloning for ComfyUI hits maintenance mode

What it does This repository is a reference implementation that brings Tencent’s IPAdapter models into ComfyUI as custom nodes. It lets you condition a diffusion generation on a single reference image, transferring either the subject’s identity or its overall style without retraining a LoRA. The author describes the effect as a “1-image lora.”

The interesting bit The project abstracts a zoo of encoders and checkpoints—CLIP-ViT variants, FaceID models, SDXL and SD1.5 weights—behind a Unified Model Loader that tries to match files by strict naming conventions. It also supports community-baked adapters like composition-only models and the Kolors pipeline.

Key highlights

• Supports style, subject, face, and composition transfer from reference images
• Bundles a Unified Model Loader that auto-detects models if you follow the exact filename conventions
• Covers SD1.5, SDXL, and Kolors model families, plus FaceID variants that need paired LoRAs
• Includes example workflows and video tutorials for masking, animation, and advanced weighting

Caveats

• Now in maintenance-only mode as of April 2025; the author will review crucial PRs but no longer plans consistent work
• FaceID models require insightface and most need a specific paired LoRA
• The Unified Loader only recognizes models if filenames match the documentation exactly; otherwise you select them manually

Verdict Worth a look if you use ComfyUI and want to clone a style or face from a reference photo without training your own LoRA. Skip it if you need active feature development or a turnkey setup that doesn’t involve wrestling with encoder downloads and exact filenames.