Microsoft has made significant strides in the domain of Generative AI Art, optimizing Direct Machine Learning (DirectML) to accelerate transformer and diffusion models such as Stable Diffusion. The aim is to leverage the Windows hardware ecosystem and enhance the AI-infused applications’ performance on it.
Stable Diffusion is an open-source model that converts text to impressive images, offering developers a platform to fine-tune their model according to specific domains, subjects, or styles. With optimized DirectML, Microsoft envisions to significantly improve Stable Diffusion’s performance across the Windows hardware ecosystem, boosting AI hardware acceleration in apps at scale. Demonstrations of what Stable Diffusion can achieve was presented in two Build sessions: “Shaping the future of work with AI” and “Deliver AI-powered experiences across cloud and edge, with Windows.”
In an effort to optimize performance, Microsoft has collaborated with the Olive team to create an optimization tool. This tool leverages DirectML to produce models specifically designed to function across the Windows ecosystem. Developers are encouraged to use Olive to convert their Stable Diffusion models to ONNX format for compatibility with DirectML, after which they can follow specific instructions on the DirectML and Olive blog to enhance performance.
Microsoft has assured developers that Stable Diffusion models with different checkpoints and/or weights, but the same architecture and layers, will work seamlessly with Olive. Examples of such models include CompVis/stable-diffusion-v1-4 and runwayml/stable-diffusion-v1-5.
To witness Stable Diffusion in action, interested parties can review the Build Breakout Session “Deliver AI-powered experiences across cloud and edge, with Windows”. Python samples for using Stable Diffusion with Olive are available, along with C++ examples of utilizing DirectML.
Microsoft has also recommended users to update their systems to the latest drivers to achieve the best performance. AMD users can expect optimized graphics drivers for AMD RDNA™ 3 devices in the next month. Intel users seeking Stable Diffusion on DirectML support can reach out to Intel Developer Relations for more information. Meanwhile, NVIDIA users with GeForce RTX 30 Series and 40 Series GPUs can experience the improvements first-hand with the GeForce Game Ready Driver 532.03.
This development signifies Microsoft’s ongoing commitment to bridging the gap between AI and traditional applications, promoting innovation and creativity in the rapidly evolving tech landscape.