up to dateinformation,SD3First invitationstestQualifications have been issued and model weights are expected to be released next month. Officials say it will be scaled up gradually and more test qualifications will be issued soon.
Previously, Stability AI released the SD3 Technical Report, revealing more details about SD3, which outperforms all current open source and commercial models in terms of typographic quality, aesthetic quality, and cue word understanding.
The report describes the new Multimodal Diffusion Transformer architecture MMDiT for SD3, which improves the system's text comprehension and spelling capabilities.
SD3 uses a corrective flow RF formulation that connects data and noise during training, enabling more direct inference paths and more efficient sampling steps.
In addition, the report describes the extended corrective flow Transformer model, which is trained at different scales by a reweighted RF formulation and the MMDiT backbone network.