SimpleTuner

A self-hosted diffusion model trainer for image, video and audio, with a web dashboard, LoRA support and distributed GPU training.

SimpleTuner is a diffusion model training toolkit for creators and researchers adapting image, video or audio models to their own datasets. Its web dashboard also supports teams sharing GPU workers.

Model fine-tuning: Train LoRA and LyCORIS adapters or full models, with support for families including Flux, HiDream, Hunyuan Video and ACE-Step.
Dataset preparation: Handle mixed image and video aspect ratios, cache processed data on disk, and generate captions on local GPUs through CaptionFlow.
Adapter comparison: Compare LoRAs through validation renders and load or save adapters in Diffusers and ComfyUI formats.
Distributed training: Spread jobs across multiple GPUs and use DeepSpeed or FSDP2 to reduce memory demands through sharding and offloading.
Shared training infrastructure: Dispatch jobs to persistent or cloud GPU workers, with priority queues, team access controls, SSO and audit logs.

The toolkit is free and open source under AGPL-3.0. Many models can train on GPUs with 16GB of memory using optimizations, while most are trainable with 24GB; requirements depend on the model. Optional S3 connections support training data stored in Cloudflare R2 or Wasabi.

It sends no data to third parties by default. External reporting, model uploads and webhooks are opt-in. Model checkpoint licenses are separate from the toolkit license, and some restrict commercial use.