AMD has released the world’s first BF16 (block FP16) Stable Diffusion 3.0 Medium model, built in collaboration with Stability AI. The model, which delivers improved image quality with lower memory usage, is now available in the latest version of Amuse 3.1 by Tensorstack. Users with compatible AMD Ryzen AI laptops can try it immediately in high-quality (HQ) mode.
Enhanced performance and efficiency on AMD XDNA 2 NPUs
The new Stable Diffusion 3.0 Medium model uses block FP16 precision and is specifically optimised for AMD’s XDNA 2 NPUs, available in Ryzen AI 300 series and Ryzen AI MAX+ laptops. It runs efficiently on machines with as little as 24GB of RAM, using only 9GB of memory to generate high-resolution images locally.
This development builds on AMD’s previous work with Stability AI on SDXL Turbo, unveiled at Computex 2024, which paired FP16 accuracy with INT8 performance. The SD 3.0 Medium model continues this focus on balancing quality and performance for generative AI applications.
A key advantage of this model is its ability to operate locally on laptops, removing the need for cloud-based processing or internet connectivity. With just a supported AMD Ryzen AI laptop and the latest Amuse 3.1 software, users can create high-quality AI-generated images without paying for subscriptions or relying on Wi-Fi.
Integrated upscaling with two-stage NPU pipeline
The SD 3.0 Medium model also features an integrated two-stage pipeline powered by the AMD XDNA 2 NPU. After generating a 2MP (1024 x 1024) image, the pipeline upscales it to a 4MP (2048 x 2048) resolution. This results in print-quality images with fine detail, making the setup suitable for both professional design work and rapid marketing asset creation.
AMD highlighted how creatives and marketers can customise AI-generated stock images tailored to their brand using the new tool. For example, users can prompt the system to generate products such as soda cans or apparel featuring their brand’s logo, allowing for quick iterations and on-demand production of brand-consistent visuals.
Prompting tips and sample outputs
According to AMD, the SD 3.0 Medium model is highly responsive to prompt structure and formatting. Users are encouraged to structure their prompts by starting with the type of image, followed by its composition and then further detail. Even minor elements such as punctuation or spacing can affect the output.
The company shared several example prompts and settings for users to recreate sample images featured in the announcement. These include a close-up shot of a toucan, a branded soda can, and fashion photography of a branded jersey. Each example uses a specific combination of prompt text, seed number, and steps, along with the “XDNA Super Resolution” setting.
The model supports negative prompts for excluding unwanted elements, though excessive use may reduce overall image quality. AMD recommends iterating prompts and running batches of 25–30 seeds to achieve the desired visual outcome.
The SD 3.0 Medium model with BF16 precision can be accessed through Amuse 3.1 by following three steps: installing the latest Adrenalin driver, installing the beta version of Amuse 3.1, and enabling “XDNA 2 Stable Diffusion Offload” while using the HQ setting.
AMD’s continued focus on NPU-optimised AI models highlights the growing demand for on-device generative AI tools, particularly those that offer efficiency without sacrificing quality. With this release, AMD positions its Ryzen AI laptops as capable tools for creators, marketers and professionals who need powerful AI generation capabilities without the limitations of cloud-based solutions.
