Stable Diffusion Runs Faster on AMD GPUs: Optimized Models Arrive | Ai image generator from image | Dall-e pretrained model | | Turtles AI

Stability AI and AMD have collaborated to optimize Stable Diffusion for AMD Radeon™ GPUs and Ryzen™ AI APUs, significantly improving image generation performance.

Key Points:

• ONNX optimized versions of Stable Diffusion available on Hugging Face.
• Up to 3.8x faster than original PyTorch models.
• Compatible with the Amuse 3.0 application for easy use.
• Support for AMD Radeon™ RX 9000 GPUs and Ryzen™ AI APUs.

Stability AI, in collaboration with AMD, has released optimized versions of Stable Diffusion models designed to take full advantage of the capabilities of AMD Radeon™ GPUs and Ryzen™ AI APUs. These models, converted to ONNX format, are available on Hugging Face with the suffix “_amdgpu” and include variants such as Stable Diffusion 3.5 Large, 3.5 Large Turbo, XL 1.0, and XL Turbo. The optimizations implemented enable up to 2.6x faster inference for SD3.5 models and up to 3.8x faster inference for SDXL models, compared to the original PyTorch versions. The Amuse 3.0 application allows users to easily use these optimized models, offering an intuitive interface and support for the advanced features of AMD GPUs. Optimizations include techniques such as flash attention, weight pruning, and custom fusions, which maximize efficiency and reduce memory usage. These improvements are especially effective on AMD Radeon™ RX 9000 GPUs, which offer 2nd generation AI accelerators with higher performance than previous generations. Additionally, Ryzen™ AI APUs benefit from efficient NPU-GPU integration, thanks to tools such as Microsoft Olive and AMD Vitis AI EP, which facilitate workload distribution and improve overall performance.

This collaboration between Stability AI and AMD represents a significant step towards optimizing generative AI models for AMD hardware, providing developers and end users with more efficient and performant tools for generating visual content.