OneTrainer

OneTrainer trains diffusion models on your own machine for people adapting image or video models to their own datasets. Its graphical interface brings dataset preparation, training and output previews into one application.

• Model training: Use full fine-tuning, LoRA or embeddings with models including Stable Diffusion, SDXL, FLUX.1, Flux.2 Dev and Klein, Qwen Image, Z-Image, Ernie Image and Hunyuan Video.
• Dataset preparation: Generate captions with BLIP, BLIP2 or WD-1.4, and create masks with ClipSeg or Rembg to focus training on selected image regions.
• Varied training samples: Combine multiple prompts per image, image augmentation and training across different aspect ratios and resolutions.
• Previews and progress: Sample the model during training, track progress in TensorBoard and use automatic backups to resume saved training state.
• Model conversion: Convert between supported model formats through the interface, with support for diffusers and ckpt models.

OneTrainer runs on Windows, macOS and Linux. A command-line interface supports headless training alongside the graphical interface. When training with EMA weights, it can optionally keep those weights in CPU memory to reduce VRAM use. The project is open source under AGPL-3.0.