Stability.AI presents the results of the Stable Diffusion 3 (SD3) test, an artificial intelligence that generates images introduced earlier. This time, they have unveiled the architecture inside and compared the test results with other models in the market.
The tests, based on human judges in three criteria – beauty, following instructions, and text in the image, revealed that SD3 outperformed almost all other models except for the beauty comparison with Ideogram 1.0.
The core architecture of SD3 is the Diffusion Transformer (DiT) that has been enhanced to be multimodal, separating text and images but using shared attention called modified multimodal diffusion transformer (MMDiT). This architecture allows the model to read text in images, with the final image containing the text according to the instructions. Stability.AI also mentioned that it can be further developed to support video generation in the future.
SD3 has three text encoders – CLIP-G/14, CLIP-L/14, and T5 XXL. Only T5 uses up to 4.7 billion parameters, which slightly affects beauty but significantly impacts text generation.
Currently, those interested in SD3 have to wait in line to use it.
Source – Stability.AI
TLDR: Stability.AI introduces the results of the SD3 test, showcasing its superior performance compared to other models in terms of generating images from text instructions. The architecture of SD3 allows it to read text in images and the model is under high demand currently.
Leave a Comment