SimpleTuner

Maintained by bghira and the Terminus Research Group, SimpleTuner trains LoRA, LyCORIS, and full fine-tunes for many generative models: image models such as SDXL, FLUX.1, FLUX.2, Qwen Image, HiDream, and Sana, video models including Wan, LTX Video, Hunyuan Video, and LongCat, and audio models like ACE-Step. It caches embeddings to disk, buckets data by aspect ratio, supports masked loss, EMA, and TREAD token dropout, and can stream datasets from S3-compatible storage. Multi-GPU and multi-node runs go through Accelerate with DeepSpeed or FSDP2 sharding, and int8, fp8, or nf4 quantization cuts memory. It runs from config files on the command line or from a web UI with a job queue, SSO, role-based access, and quotas for shared GPU clusters. It installs with pip and runs on NVIDIA cards from the RTX 3080 class up, AMD 7900 XTX or MI300X, and Apple M3 Max machines with 24 GB or more; LoRA on 12B-plus models wants 24 GB of VRAM or more. AGPL-3.0 permits commercial use but requires releasing source for modified versions offered over a network, and each base model's license still governs the trained weights.