finetrainers

finetrainers is a training toolkit from the Hugging Face Diffusers ecosystem focused on fine-tuning video and image generation models. Supported models include LTX-Video, HunyuanVideo, CogVideoX, and Wan for text-to-video, plus CogView4 and Flux for text-to-image, with LoRA, full-rank fine-tuning, and conditional control training all available. The library treats memory efficiency and scale as equal goals: it implements DDP, FSDP2, HSDP, and context parallelism for multi-GPU training while offering gradient checkpointing, precomputation of embeddings, FP8 weight casting, and multiple attention backends (flash, flex, sage, xformers) so a single consumer GPU can still train models; an LTX-Video LoRA fits in about 5 GB of VRAM. Common dataset formats are auto-detected and multi-resolution bucketing is built in. Apache 2.0 licensed and under active development, it is used by researchers and practitioners customizing open video models for domains such as visual effects and robotics.