Stability AI releases open weights for Stable Diffusion 3 Medium

Background

The Medium version is a 2-billion-parameter member of the Stable Diffusion 3 family, and this release made its weights downloadable.

What happened

On 12 June 2024 Stability AI released the weights of Stable Diffusion 3 Medium. ^[1] The weights were published on Hugging Face, and the release included a TensorRT-optimised version of the model. ^[1]

A 2-billion-parameter model is small enough to run on a personal computer with a suitable graphics card, while larger models in a family typically need more memory. Downloadable weights can be run locally, without relying on a hosted service. A non-commercial license permits use for research and personal purposes, but paid products or services normally need a separate agreement.

Hosted services and local downloads serve different users. A hosted service runs the model on the provider's servers, so the user needs only a browser or an app, while local weights require the user to manage the software and the hardware. Offering both routes lets people with different equipment and needs choose the one that suits them.

Key details

Stable Diffusion 3 Medium is a 2-billion-parameter diffusion transformer with a 16-channel VAE. A VAE, or variational autoencoder, is the component that compresses images into the smaller representation the model works in and then decodes the result back into pixels. ^[1]

TensorRT is NVIDIA's library for speeding up neural networks on its GPUs, so a TensorRT-optimised version is one prepared to run faster on that hardware. ^[1]

Stability said that its small size suits consumer PCs and laptops. ^[1]

The release was under a non-commercial license, with a separate Creator License for professional use. ^[2]

The model was also available through the Stability Platform API, Stable Assistant and Stable Artisan. ^[1]

Impact

The release made the design usable on ordinary personal computers and laptops, while keeping commercial professional use under a separate license. This split between a free non-commercial tier and a paid professional tier is the arrangement the company used for this model. ^[1][2]