OmniChar: One Character Across Every Model, Carried in a Portable .char File
OmniChar is a GPL-3.0 open-source, fully local character-consistency tool that has been drawing attention overseas. It packages face and wardrobe into a tiny portable .char file: build a character once and it stays itself — same face, same body, same outfit — across every model you use, solving the difficulties LoRA-based approaches face with off-angle depictions and feature drift. Technically, a .char stores an SFace face signature and a DINOv2 subject signature extracted from your photos; FLUX.2 has a native multi-reference channel and reads those references directly with nothing to train, while Krea 2 has no reference channel and instead takes a LoRA trained from the same photos on the built-in Trainer canvas. One .char file carries both forms, so switching models does not mean rebuilding the character, and every file records how it was made and who it belongs to. It runs on Inline Core, a from-scratch diffusion engine that serves the web UI, runs models, keeps the project database and drives the timeline as a single Python process, supporting Z-Image Turbo, FLUX.2 and Krea 2 for images plus MiniMax H3 and LTX-2.5 for video, on NVIDIA, AMD or CPU with Python 3.11+. It is also available as ComfyUI nodes and can reach hosted Seedance and Kling through API nodes. Community reception has been largely positive, many welcoming it as a way to streamline character management, while others suggest using it alongside existing LoRA and reference-image workflows depending on the scenario.
