The Stable Diffusion 3 model introduces the DiT (Diffusion Transformer) architecture, which significantly improves quality. Stable Diffusion 3 outperforms other text-to-image generation systems. The parameter count of Stable Diffusion 3 ranges from 800M to 8B. The SD3 architecture is developed in collaboration with core members of Sora and an assistant professor from New York University.

The technical report on Stable Diffusion 3 reveals that SD3 adopts the Multimodal Diffusion Transformer architecture MMDiT. SD3 introduces reweighted flow technology to enhance performance. The report discusses the scaling research of SD3 and anticipates future performance improvements along with issues and suggestions regarding the text encoder.

Stable Diffusion 3 is the most powerful text-to-image model. It adopts the MMDiT architecture, demonstrating performance that surpasses existing text-to-image generation systems. Stable Diffusion 3 excels in visual aesthetics, text adherence, and layout compared to other advanced models. The MMDiT architecture combines DiT and rectangular flow forms, processing image and language representations through independent weight sets. Stable Diffusion 3 offers flexibility.