AMD and Stability AI have unveiled a significant collaboration to adapt Stable Diffusion 3.0 Medium for Stability Amuse, a creative Artificial Intelligence art platform. By refining the model architecture to operate efficiently on Neural Processing Units (NPUs), the partnership enables enhanced image generation and improved text processing directly on devices equipped with AMD´s latest Ryzen AI XDNA 2 NPUs. This development builds on technology showcased at Computex 2024, where AMD and Stability AI introduced SDXL Turbo, the world´s first block FP16 stable diffusion model, mixing FP16 accuracy with performance closer to INT8 compute levels.
The new block FP16 SD 3 Medium model delivers superior picture quality with considerably reduced memory requirements compared to previous iterations, functioning smoothly on laptops with 24 GB RAM and consuming just 9 GB during operation. This optimization is crucial, as it makes high-precision, local Artificial Intelligence image generation feasible on mainstream hardware, sidestepping the heavy demands of quantization. A two-stage processing pipeline leverages the XDNA 2 NPU to boost output from 2 megapixels (1024 x 1024) up to 4 megapixels (2048 x 2048), pushing desktop-grade creative output onto portable devices.
Originally, Amuse required GPU acceleration for tasks involving Stable Diffusion Medium, restricting its accessibility to users with dedicated GPUs. Now, the latest Amuse 3.1 release lets users select between GPU and NPU acceleration, broadening hardware compatibility and enabling efficient, local Artificial Intelligence workflows on Ryzen AI-powered laptops. To access this feature, users need to install the latest AMD Adrenalin Driver and Amuse 3.1 Beta, enable HQ mode in EZ Mode, and activate the XDNA 2 Stable Diffusion Offload option. This marks a pivotal step for creative professionals and hobbyists seeking high-resolution image generation without reliance on cloud processing or high-end GPUs.
