One character file, every diffusion model

What it does

OmniChar (branded “Inline Studio” in the app) is a GPLv3 Python studio for AI character generation. You build a character once — face, body, wardrobe — and carry it across shots and models, generating locally on your own GPU or through hosted API models. It also trains LoRAs in-app and keeps every render as a versioned take, which is the part most character tools quietly skip.

The interesting bit

The .char format is the pitch: one portable file that works across every supported model, instead of a folder of prompts and a prayer. The README is also unusually honest about hardware — it distinguishes measured benchmarks from interpolated ones and admits which code paths (Apple Silicon, AMD ROCm) nobody has actually run.

Key highlights

• Trains and generates with FLUX.2, FLUX.1, Krea 2, Z-Image Turbo, and video models MiniMax H3 and LTX-2.5
• LoRA training fits on a 16GB card for most models (FLUX.2 klein at ~8.6GB); a 512px-trained LoRA applies at any generation resolution
• Multi-GPU support splits a single image’s denoise across cards, automatic with 2+ GPUs
• Runs fully offline for local generation; hosted API models available for the GPU-less
• Installer reads your NVIDIA card’s compute capability and picks the right CUDA build, RTX 50-series included

Caveats

• Generation is much hungrier than training: Krea 2 inference peaks near 36GB at 1024px, so a 40GB card is the practical floor there
• Apple Silicon and AMD ROCm paths exist but are untested; LTX-2.5 training wants 48GB
• The README is install-heavy and light on how the .char format actually works under the hood — the spec itself stays vague

Verdict

Worth a look if you make AI film or comics and are tired of characters drifting between shots. Skip it if you’re on a Mac or AMD card — those paths are untested — or if you just need one-off images.