OneTrainer

Nerogar's OneTrainer gathers the chores of training image and video diffusion models into one desktop app: captioning, mask making, format conversion, preview sampling, and the training loop itself. Supported families include Stable Diffusion 1.5 through 3.5 and SDXL, FLUX.1 and Flux.2, Chroma, Qwen Image, Z-Image, PixArt, Sana, Stable Cascade, and Hunyuan Video. You can run full fine-tunes, LoRA adapters, or textual inversion embeddings, and masked training limits the loss to chosen image regions. Extras include aspect ratio bucketing, mixed resolutions in one run, EMA weights parked in system RAM to spare VRAM, augmentations, backups for resuming, and TensorBoard logging. Captions come from BLIP, BLIP2, or WD-1.4 and masks from ClipSeg or Rembg. Setup is a git clone plus install.bat or install.sh on Python 3.10 to 3.13, with requirement sets for NVIDIA CUDA and AMD ROCm; a Tk interface then opens, and each feature also runs as a headless script. The AGPL-3.0 license permits commercial training, but network-served modified versions must publish source. Good for hobbyists and studios making custom LoRAs.