Stable Diffusion Now Optimized for AMD Radeon™ GPUs and Ryzen™ AI APUs

---

Stable Diffusion, the popular generative model family, has received a significant performance boost on AMD hardware. In collaboration with AMD, Stability AI has released ONNX-optimized versions of several key models, specifically tailored to run faster and more efficiently on AMD Radeon™ GPUs and Ryzen™ AI APUs. This update is crucial for developers and creators looking to leverage the latest advancements in generative AI without sacrificing speed or quality.

Key Technical Changes

• ONNX Optimization: The new models are optimized using ONNX (Open Neural Network Exchange), a format that allows models to be more portable across different platforms and hardware.
• AMD Hardware Support: These optimizations target AMD Radeon™ GPUs and Ryzen™ AI APUs, ensuring better performance on these specific architectures.

Available Models

The following models have been optimized for AMD hardware:

Stable Diffusion 3.5 (SD3.5) Series:

• Stable Diffusion 3.5 Large:
  • Hugging Face Link
• Stable Diffusion 3.5 Large Turbo:
  • Hugging Face Link

Stable Diffusion XL (SDXL) Series:

• Stable Diffusion XL 1.0:
  • Hugging Face Link
• Stable Diffusion XL Turbo:
  • Hugging Face Link

Performance Improvements

The AMD optimizations yield significant speedups compared to the base PyTorch models:

• SD3.5 Models: Up to 2.6x faster inference
• SDXL Models: Up to 3.8x faster inference

These performance gains are crucial for real-world applications, such as content creation and real-time processing, where speed is a critical factor.

Implementation Details

• ONNX Runtime Integration: The optimized models integrate seamlessly with ONNX Runtime, making it easy to drop them into existing workflows.
• Model Quality: Despite the optimizations, the quality of the model outputs remains unchanged, ensuring that users can expect the same high standards as before.
• Compatibility: These models are compatible with any environment that supports ONNX Runtime, including popular frameworks like TensorFlow and PyTorch.

Why It Matters

Generative AI is becoming increasingly prevalent in various industries, from media and entertainment to scientific research. Optimizing these models for leading hardware platforms ensures that developers can deploy them efficiently, reducing latency and improving user experience. This collaboration between Stability AI and AMD is a step forward in making advanced AI more accessible and practical.

How to Try It Out

End users can try out the AMD-optimized models using Amuse 3.0, an AMD tool designed for testing and deploying optimized models. You can find more details on how to get started on the AMD Community Forum.

For a deeper dive into the technical aspects of these optimizations, check out AMD’s detailed blog post on GPUOpen.

Conclusion

The release of AMD-optimized Stable Diffusion models is a significant milestone in the ongoing effort to make generative AI more accessible and efficient. Whether you’re a developer looking to integrate these models into your projects or a creator exploring new creative possibilities, this update brings faster performance and better hardware support to the table.