One GUI to train every diffusion model you own

What it does

OneTrainer is a training workbench for diffusion models: full fine-tuning, LoRA, and embeddings across everything from Stable Diffusion 1.5 through SDXL, FLUX.1 and Flux.2, Chroma, PixArt, Sana, and even Hunyuan Video. It handles both diffusers and ckpt formats, and it doesn’t stop at training — it also captions your dataset (BLIP, BLIP2, WD-1.4), generates masks (ClipSeg, Rembg), converts model formats, and samples images mid-training without leaving the app.

The interesting bit

The boring parts are where the value hides: automatic backups that let you resume training exactly where it died, aspect-ratio bucketing so mixed-resolution datasets don’t fight the trainer, and optional EMA weights parked in CPU RAM to save VRAM. It also implements noise scheduler rescaling from an actual arXiv paper, which is more homework than most training UIs did.

Key highlights

• Full fine-tuning, LoRA, and embeddings across a long list of models, including video (Hunyuan) and inpainting variants
• Masked training to focus learning on specific regions of samples
• Dataset tooling: auto-captioning, mask generation, per-image loss calculation
• GUI and CLI modes — every function exists as a standalone script in scripts/
• TensorBoard integration and a sampling UI that works while training runs

Caveats

• Requires Python between 3.10 and 3.14, and some Linux distros need extra system packages (libGL, tkinter)
• Bug reports must include a generated debug_report.log or they’ll likely be closed — a firm but fair policy

Verdict

If you fine-tune diffusion models regularly and are tired of stitching together captioning scripts, conversion tools, and training configs, this is your workbench. If you train one model once a year, it’s a lot of app for one job.