OmniChar: A Consistent AI Character in a Single .char File

The open-source project OmniChar (GPLv3) introduced a portable .char format that stores a consistent AI character — face, body, clothing, and a trained LoRA adapter — and transfers it between FLUX.2, Krea 2, MiniMax H3, and Z-Image Turbo, measuring appearance drift with a number.

Compiled by Sergey KostenchukPublished 2026-10-03Updated 2026-10-03

2026-10-03 Coding NVIDIA

🛠 OmniChar: A Consistent AI Character in a Single .char File

The open-source studio OmniChar (GPLv3) makes AI characters portable: a .char file stores the face, body, clothing, and LoRA adapter of a single hero. Each render receives a continuity score of 0–100 — appearance drift is visible as a number. FLUX.2, Krea 2, MiniMax H3, and Z-Image Turbo are supported, with a ComfyUI node and a cloud beta available.

🌍 A hero's identity is transferred with a single .char file, compatible with ComfyUI and the cloud — instead of hours of LoRA training or face adapters that transfer only the face. For MiniMax H3 inference, ~40 GB of weights and 64 GB of RAM are required.

👤 With an NVIDIA card (a 16 GB path is tested) or a fal account, a hero is set up in seconds: git clone, ./webui.sh --install --extra all, web UI at 127.0.0.1:8848. However, multi-character scenes break, and 5 references at 1024px slow down the step by ~6 times.

Source 1: https://github.com/omnichar/OmniChar Source 2: https://www.omnichar.org/characters/how-it-works

Sources