ZenTrainer 0.4 trains image and video LoRAs inside ComfyUI, with live previews, on NVIDIA or AMD
ZenTrainer is a free, open-source app that trains LoRAs inside ComfyUI, using the models and GPU it already has, and shows samples and loss while it learns. Version 0.4.0, released on 30 October, trains ten models end to end, including MOSS Image, Krea 2, FLUX 3, Qwen Image 2.1, SDXL and MiniMax H3 video.
What it does
Install it from ComfyUI Manager and click the lotus in the top bar. In Simple mode you pick a recipe, add 15 to 40 photos and name the run; missing models download and the photos are captioned for you, and the LoRA lands in output/zentrainer/loras/, ready for ComfyUI's LoRA loaders. Each run is one page with samples denoising beside the loss: you can pause, sample, save or change the learning rate mid-run, resume from any checkpoint, see validation loss and a likeness score for character LoRAs, and try a prompt on any checkpoint. Datasets are folders in ComfyUI's input, with auto-captions from Qwen-VL or Gemma, bulk caption edits and SAM 3 masks so only the subject trains.
Worth knowing
New in 0.4.0: Anima, FLUX.2 Klein and SDXL with its finetunes (Pony, Illustrious), recipes stored as plain JSON files, and LoRA metadata that no longer carries your dataset folder names or captions. Needs ComfyUI 0.41 or newer on Linux or Windows with NVIDIA (CUDA) or AMD (ROCm); macOS is not supported yet. It can also run with its own pinned ComfyUI so training can't break your main install, plus a zt command line and a Docker image for rented GPUs. MIT licence; the likeness score's InsightFace models are non-commercial, with fal's AuraFace as the commercial option.