Stable Diffusion 3: the paper reveals the MMDiT architecture and rectified flow
In one sentence On March 5, 2024 Stability AI publishes the Stable Diffusion 3 paper (announced on February 22): MMDiT transformer with separate weights for text and image, rectified flow training, variants from 800M to 8B parameters.
Stable Diffusion 3 swaps the engine of previous models: instead of the classic UNet it uses a transformer, the same kind of architecture as language models. Stability AI announced it on February 22, 2024 with a waitlist for the preview; on March 5 it publishes the paper with the technical details.
According to human preference tests run by Stability, SD3 matches or beats DALL·E 3, Midjourney v6 and Ideogram v1 on prompt following, text rendering in images and aesthetics.
Weights only arrive on June 12, 2024, with SD3 Medium.
Companies
Stability AI
Tools
Stable Diffusion 3, SD3
Tags
Sources