open-mmlab/mmagic

MMagic (formerly MMEditing) is a pillar of the OpenMMLab ecosystem, which is the de facto standard for academic computer vision research. With over 7,400 stars and 1,000+ forks, it possesses significant community momentum and 'data gravity' via its massive model zoo (70+ algorithms including Stable Diffusion, ControlNet, and Real-ESRGAN). The defensibility stems from the 'MM' ecosystem lock-in: once a research lab adopts MMEngine and MMCV, the switching cost to move away from MMagic to a fragmented set of individual repos is high. Its main competitor is Hugging Face's 'diffusers' library; while Hugging Face dominates the generative weights distribution, MMagic excels in modularity for researchers who need to hack model architectures or combine generation with classical restoration tasks (SR, inpainting). Frontier labs (OpenAI/Google) pose a medium risk because they focus on releasing weights/APIs rather than the underlying training frameworks for the community. The displacement horizon is long because the project is backed by a structured academic community (MMLab at CUHK) that consistently integrates new SOTA papers faster than commercial entities.