# OneTrainer

A one-stop, UI-driven tool for training and fine-tuning diffusion image and video models, supporting full fine-tuning, LoRA and embeddings across many model families and formats.

OneTrainer is a one-stop solution for diffusion model training. It supports a wide range of models, including Ernie Image, Z-Image, Qwen Image, FLUX.1, Stable Diffusion 1.5 through 3.5, SDXL, Würstchen, Stable Cascade, PixArt-Alpha and Sigma, Sana, Hunyuan Video and inpainting models, in both diffusers and checkpoint formats, with full fine-tuning, LoRA and embedding training methods.

It can mask training to focus on specific parts of sample images, augment a dataset with random rotation, brightness, contrast or saturation transforms, train on multiple prompts per image and multiple resolutions at once with automatic aspect-ratio bucketing, rescale the noise scheduler to correct flawed diffusion noise schedules, and train an EMA model with optional CPU-side weight storage to reduce VRAM use. Dataset tooling can auto-caption images with BLIP, BLIP2 or WD-1.4, or generate masks with ClipSeg or Rembg, and a TensorBoard integration tracks training progress, with a sampling UI for testing the model mid-training.

It installs from source with an automatic or manual installer, requires Python 3.10 or newer, and is released under the AGPL-3.0 license.

Website: https://github.com/Nerogar/OneTrainer