DiffSynth-Studio is ModelScope’s open diffusion engine: a single framework for running and training state-of-the-art image and video generation models such as FLUX, Qwen-Image and the Wan video family. Built by the team behind China’s largest model hub, it has earned about 13,300 stars on GitHub and is updated continuously through 2026.
Every new diffusion model usually arrives with its own incompatible codebase, which makes experimentation and fine-tuning painful. DiffSynth-Studio unifies many architectures behind shared pipelines for inference and training, so the same workflow loads a text-to-image model, a video model or an editing model, and the same training code can fully fine-tune a model or train a LoRA. The framework pays unusual attention to efficiency: layered offloading and other memory optimizations let large models train and run on consumer-grade GPUs that would otherwise be too small. It plugs directly into the ModelScope model zoo for weights, and the companion DiffSynth Academy materials teach the underlying techniques. For researchers, that combination makes it a convenient lab for studying how modern diffusion models actually work; for creators, it is a practical way to fine-tune current models on their own data.
DiffSynth-Studio is free and open source under the Apache-2.0 license. Models and compute are your own; there is no paid tier for the framework itself.
DiffSynth-Studio is for creators and researchers who fine-tune or deploy diffusion models and want one maintained codebase covering many current architectures, instead of juggling a different repository for every new model release.