AMD has launched the first BF16 Stable Diffusion 3.0 Medium model for local image generation, optimised for AMD XDNA 2 NPUs and available within the Amuse 3.1 release.
The new model, created in collaboration with Stability AI, supports high-quality image generation while reducing memory usage, enabling operation on laptops with as little as 24GB of RAM by consuming just 9GB for execution. This development means that users can run advanced AI image tasks on consumer devices equipped with Ryzen AI 300 Series or Ryzen AI Max processors with more than 50 NPU TOPS.
Technical details
According to AMD, the BF16 precision format, also referenced as block FP16, blends the accuracy characteristic of FP16 with the computational performance of INT8. This combination supports more sophisticated generative AI workloads while maintaining efficiency within hardware capabilities. The Amuse 3.1 software release, which enables access to this model, further extends these benefits to compatible consumer hardware.
The model features a 2-stage processing pipeline utilising the XDNA 2 NPU, upscaling initial 2MP (1024×1024) images to a 4MP (2048×2048) resolution. This enhancement, as stated, is intended to deliver “print quality images tailored to your specifications on-the-go.”
AMD provided further details about practical usage and potential applications, referencing capabilities for graphics design and rapid creation of custom marketing assets. These tasks can be accomplished offline, without the need for a persistent Wi-Fi connection or recurring subscription fees for image generation.
Usage and compatibility
The new BF16 SD 3.0 Medium model is designed to function on systems featuring Ryzen AI 300 series or Ryzen AI Max processors, each paired with an AMD XDNA 2 NPU delivering at least 50 TOPS of performance. AMD highlighted that this enables high precision AI image generation to be accessible on a wider range of laptop devices than previously possible, without requiring machines equipped with 32GB or more RAM.
AMD explained how users can access the new capabilities: “Try SD 3.0 Medium (with BF16 precision) right now on an AMD Ryzen AI 300 series or Ryzen AI MAX+ laptop with at least 24GB memory right now by following three simple steps: SD3 Medium (NPU) Hardware Requirements: AMD Ryzen AI 300 series or Ryzen AI MAX+ laptop equipped with a 50 TOPs or higher AMD XDNA 2 NPU and at least 24GB of system RAM. Download and Install the latest Adrenalin Driver. Download and Install Amuse 3.1 Beta. In EZ Mode, move the slider all the way to HQ and toggle ‘XDNA 2 Stable Diffusion Offload’.”
Image generation and prompting
AMD noted that prompt engineering remains crucial for optimal results with Stable Diffusion 3.0 Medium. They advised: “Stable Diffusion 3.0 Medium is an extremely capable model that is very sensitive to the prompt content, structure and order. Here are some prompting tips in order to get the best quality: The same size, seed, steps, sampler, scheduler and model combo should yield the same image. You generally want to start by describing the type of image then the structural components of an image and then transition to details and other context. Not every seed will yield a perfect result. Typically, you want to iterate your prompt till you get to the visual structure you want (typos are fine) and then automate a batch with 25-30 seeds. Even spaces and full stops make a difference. You can utilize negative prompts to remove elements you don’t want from an image but doing this excessively will have a quality impact on the image.”
The company also supplied sample prompts, seeds, and settings that allow users to recreate test images, including detailed instructions for users aiming to replicate specific outputs.
Further guidance and considerations
AMD reminded users that Amuse 3.1 is classified as beta software, supplied by a third-party provider, and may show instability or bugs. They also commented on applicable licencing: “Image generation using SD 3 Medium is regulated by the Stability AI Community Licence and is free for personal use and for SMEs under $1 million in annual revenue. Licensing requirements may change at the sole discretion of the third party. For complete licensing requirements please refer to: LICENSE.md · stabilityai/stable-diffusion-3-medium at main. Internet connection is required to download the model and other configuration files.”
The company clarified feature availability based on processor compatibility, noting that AMD Ryzen AI includes a combination of an AI engine, Radeon graphics engine, and Ryzen processor cores subject to enablement from both OEM and ISV partners. They advised customers to check system compatibility prior to purchase.
At Computex 2024, AMD introduced the world’s first block FP16 stable diffusion model: the SDXL Turbo. The breakthrough model combined the accuracy of FP16 with the performance of INT8 and was a collaboration between AMD and Stability AI.
The BF16-enabled SD 3.0 Medium model is intended to address the expanding demand for on-device AI capability while managing the memory constraints present in most mainstream laptops, and is available immediately with the release of Amuse 3.1 for supported hardware.