AMD has just unveiled SD 3 Medium, the world’s first BF16 (block FP16) Stable Diffusion model optimized for its new XDNA 2 AI NPUs.
This breakthrough not only boosts image generation speed but also slashes memory usage, making it ideal for high-performance AI tasks on laptops with limited RAM.
Debuted at Computex 2024, this new model builds on the earlier SDXL Turbo (developed with Stability AI), combining the precision of FP16 with the efficiency of INT8. Now, the SD 3 Medium model takes things further, providing sharper images and faster results thanks to AMD’s collaboration with Stability AI and integration into Ryzen AI laptops.
Users can start generating high-quality images locally using Amuse 3.1 Beta by Tensorstack. Just switch on “XDNA 2 Stable Diffusion Offload” in HQ mode to experience the power of the SD 3 Medium model. And you don’t need a workstation to do it-just a Ryzen AI 300 or Ryzen AI MAX+ series laptop with a 50 TOPs+ NPU and at least 24GB of RAM.
The model is impressively optimized, requiring just 9GB of memory to produce stunning AI images. AMD also implemented a 2-stage pipeline that upscales output from 2MP (1024×1024) to a crisp 4MP (2048×2048), perfect for design or print-ready assets. This makes it a game-changer for creators and marketers who need customizable visuals on the fly.
With an AMD Ryzen AI-powered device, users can rapidly prototype and refine branded content anywhere, even mid-flight. The XDNA 2 NPU delivers robust AI processing while keeping power usage in check, giving creators freedom and performance without compromise.