VibeusZenTrainer brings LoRA training into ComfyUI: choose a model, add... - Vibeus
ZenTrainer brings LoRA training into ComfyUI: choose a model, add your images and let the interface fill in model files, captions and settings. The 0.4.0 release is free and open source under MIT, with live training previews and a likeness score for character checkpoints. It reuses the models already stored in your ComfyUI folders.
The developer reports completed training runs with MOSS Image, Krea 2, Fibo 2.0, FLUX 3 Image, FLUX.2 Klein 9B, Qwen Image 2.1, Z-Image, Anima, SDXL and its finetunes, including Pony and Illustrious, plus MiniMax H3 video. Other models detected through ComfyUI are labelled untested. NVIDIA and AMD/ROCm are supported on Linux and Windows; macOS support has not arrived.
Install through ComfyUI Manager, restart and open the lotus button in the top bar. A separate installation with a pinned ComfyUI version can keep training apart from changes to your main setup. The current README requires ComfyUI 0.41 or newer. Simple mode starts from a recipe and 15–40 photos; Advanced exposes the individual settings. Runs show previews beside loss and validation measurements, can resume from checkpoints and can queue sequentially.
The attached complete trailer shows image captioning, live samples and MiniMax H3 motion. Its two community pixel-art LoRAs—4Bit Alchemy by Addictedd and Nostapixx by JadedScorpion—appear with their creators’ permission. One user reports successful training across several models and requested 256-pixel resolution, which the developer subsequently added. Another reports an unresolved H3 training-adapter error. Memory needs depend on the chosen model; the thread does not establish a universal minimum.
Source code and the trailer link are in the original post. The repository already lists 0.4.1, a Docker-build fix. Model licences remain separate from the MIT code licence; in particular, the README identifies InsightFace face models as non-commercial and describes AuraFace with YuNet as the commercial-use alternative.