Stable Diffusion 3 (early preview)
Stable Diffusion 3 was previewed as a diffusion transformer trained with flow matching, at 800M to 8B parameters and waitlist-only.
Replaces the U-Net with a multimodal diffusion transformer that keeps separate weights for text and image tokens and trains with rectified flow. The paper (2024-03-05) reports predictable scaling with validation loss and better typography and human-preference scores.
- Date
- Thursday, 22 February 2024
- Lab
- Stability AI
- Kind
- model
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Parameter range | 800M-8B model suite at announcement | company |
The MMDiT + rectified-flow recipe was later adopted by Flux (same founders) and many successors. The paper is titled Scaling Rectified Flow Transformers for High-Resolution Image Synthesis.
Sources
- stability.ai/news-updates/stable-diffusion-3
- arxiv.org/abs/2403.03206
- techcrunch.com/2024/02/22/stable-diffusion-3-arrives-to-solidify-early-lead-in-ai-imagery-
This record was checked against its sources on 6 October 2026. How we check