Musubi Tuner

Musubi Tuner is kohya-ss's answer to training LoRAs for the current generation of video and image diffusion transformers on consumer hardware. From the author of the sd-scripts trainer that dominated the Stable Diffusion era, it supports HunyuanVideo, Wan 2.1 and 2.2, FramePack, FLUX.1 Kontext, FLUX.2 dev and klein, Qwen-Image, Z-Image, Kandinsky 5, and other recent architectures, with memory-saving techniques such as latent and text-encoder output precaching and block swapping between GPU and system RAM. The documented baseline is 12 GB of VRAM or more for image training and 24 GB or more for video work, with 64 GB of main memory recommended. Install is pip based on Python 3.10 through 3.12 with PyTorch 2.5.1 or later, plus an experimental uv path, and training runs from command line scripts driven by TOML dataset configs. Repository docs cover each architecture and advanced settings. Most of the code carries the Apache-2.0 license, with the HunyuanVideo derived directories following Tencent's licenses instead, and the repository was updated continuously through mid 2026 as a standard trainer for open video model LoRAs.