DiffSynth-Studio
The ModelScope community's open-source diffusion engine for running and training image, video and audio models on modest GPUs.
DiffSynth-Studio is an open-source diffusion model engine built and maintained by the ModelScope community team. It integrates mainstream open diffusion models for image, video and audio generation, along with image-quality metric models. Its README highlights four features: VRAM management that moves parameters between disk, memory and VRAM so low-memory consumer GPUs can run large models, quantization to precisions such as NF4 and INT8, training support for almost every model that can do inference, including LoRAs and adapters, and a split-training mode that divides training into two stages.
The project adds new models quickly. Its news shows support for FLUX.3 Image Klein on 16 October 2026, MOSS-Image on 20 October, FIBO 2.0 on 21 October and MiniMax-H3-Image on 10 November, each with inference and training and, for most of them, low-VRAM inference. It also released EntroPack, a PyTorch tensor compression library that serves as a quantization backend, and the README says its 3.5-bit MOSS-Image build can run with as little as 4 GB of VRAM.
It is released under the Apache 2.0 license and installs from source or from PyPI with pip. According to its installation guide it supports NVIDIA GPUs, AMD GPUs through ROCm and Ascend NPUs. Models download from ModelScope by default, and an environment variable switches to the international ModelScope site or to Hugging Face. It is code, not an app, although the same team also publishes a ComfyUI node package and a lightweight LoRA training tool built on it.
Good for
- Running large image and video diffusion models on GPUs with limited VRAM
- Training LoRAs or full models for a wide range of recent open image and video models
- Researchers and developers who want a code-level engine rather than a point-and-click app
Watch out for
- The README says the team is small, so new features and answers to issues can be slow
- Model downloads default to ModelScope, and model IDs can differ between ModelScope and Hugging Face
- Major version updates retired some older features, so old tutorials may not match the current code