Back to Gap Map Model components / Image, video, 3D & music generation

DiffSynth-Studio

open source / Overall score: 3.2

DiffSynth-Studio is the ModelScope team's Python engine for running and training diffusion models. It supports inference and full or LoRA training for image models (FLUX, Qwen-Image, Z-Image and others), video models (Wan, LTX-2, MiniMax H3) and audio models (ACE-Step, YuE2, MiniMax-Music3), and powers ModelScope's hosted generation service.

The PyPI package diffsynth is published by "ModelScope Team", carries the repository README as its description, and is the package the README's PyPI badge points to; its metadata names no repository URL. The README recommends a source install. A third engine beside stable-diffusion.cpp and Nunchaku, and the only one here that trains as well as runs models.

Openness

5 high confidence

5.0

license: Apache-2.0(OSI
source: public(modelscope/DiffSynth-Studio)
core features withheld: no — no ee, enterprise or proprietary path

DiffSynth-Studio is released under Apache 2.0 with its full source public. ModelScope runs a hosted service on it but sells no separate edition of the engine.

https://ungh.cc/repos/modelscope/DiffSynth-Studio/files/main recorded 2026-09-27
Full file list of 1568 files: no ee/, enterprise/ or proprietary/ path.
https://ungh.cc/repos/modelscope/DiffSynth-Studio/files/main/LICENSE recorded 2026-09-27
LICENSE body: "Apache License / Version 2.0, January 2004".
https://ungh.cc/repos/modelscope/DiffSynth-Studio/files/main/README.md recorded 2026-09-27
README: "ModelScope AIGC Zone (for Chinese users): Productized features powered by DiffSynth-Studio as the core inference and training engine"; no paid edition of the engine is named.

Adoption

2 medium confidence

2.0

Adoption is measured as PyPI downloads of the diffsynth package. The README recommends installing from source, which this figure does not count, and the monthly figure sits just over the edge of its range, so the next read may place it a level lower.

https://pypi.org/pypi/diffsynth/json recorded 2026-09-27
PyPI diffsynth 2.1.8: author "ModelScope Team", summary "Enjoy the magic of Diffusion models!", the repository description, license_expression Apache-2.0.
https://pypistats.org/api/packages/diffsynth/recent recorded 2026-09-27
pypistats for diffsynth: last_month 10736, last_week 1562.

Capability

4 high confidence

4.0

DiffSynth-Studio runs image, video and music models, and can train almost every model it runs. It sits a step below ComfyUI because it runs no 3D generation model.

https://ungh.cc/repos/modelscope/DiffSynth-Studio/files/main/README.md recorded 2026-09-27
README: "covering image generation, video generation, audio generation, and image quality metrics models" and "Almost every model that supports inference also supports training"; the supported-model table lists Wan, LTX-2, Qwen-Image, MiniMax-H3, FLUX.1, FLUX.2, ACE-Step, Z-Image, YuE2 and MiniMax-Music3, and no 3D model.

Verified 2026-09-27