finetrainers

Why we included this project

Fine-tuning a modern video or image diffusion model usually means wrestling with VRAM limits, but this library tackles that head-on. Built on the Diffusers stack, it supports supervised fine-tuning and LoRA for models like CogVideoX, LTX-Video, HunyuanVideo, Wan, Flux, and CogView4, and its FP8 weight casting and gradient checkpointing can squeeze some video LoRA runs into roughly 5 GB of VRAM. Training engineers get distributed training across DDP, FSDP-2, and context parallelism, plus a standardized way to define custom model architectures and automatic detection of common dataset layouts. Ready-to-run example scripts and dataset tooling for curating small, high-quality video sets mean you can move from raw clips to a real training run without assembling the plumbing yourself. One caveat: the main branch is actively developed, so pin a stable release tag if you want reproducible results.