Stability AI has released a smaller yet highly capable version of its flagship model, Stable Diffusion 3, dubbed Stable Diffusion Medium. This new model is designed to run efficiently on consumer-grade hardware, making it an attractive option for users and organizations with resource constraints. With only 2 billion parameters, compared to the 8 billion parameters of the initial Stable Diffusion 3 Large, the Medium model is capable of running on hardware with as little as 5GB of GPU VRAM. Despite its smaller size, the model boasts an exceptionally high level of quality, comparable to its larger counterpart, with capabilities including photorealism, prompt adherence, typography, resource efficiency, and fine-tuning.
In an interview, Christian Laforte, co-CEO at Stability AI, emphasized the model’s ability to deliver highly realistic image outputs, thanks to its 16-channel VAE. He also highlighted its remarkable prompt understanding, spatial awareness, and adaptability. The model’s resource efficiency makes it an ideal choice for environments where resource management is critical. This release marks a significant shift in the paradigm of generative AI workloads, moving away from reliance on beefy Nvidia GPUs.
