What changed
AMD announced TeaCache, a technique that accelerates Stable Diffusion 3 and 3.5 image generation on AMD Ryzen AI neural processing units (NPUs) by reusing intermediate outputs across the denoising loop. This optimization avoids redundant computation without requiring model retraining or architectural changes.
What this means in practice
TeaCache works by measuring the relative L1 distance between intermediate representations across timesteps. If the change is sufficiently small, it reuses cached outputs instead of recomputing them. This approach reduces latency while maintaining image quality. The technique is integrated into an ONNX-based pipeline, allowing developers to apply caching strategies like naive caching or rescaled caching. AMD tested the method on Ryzen AI hardware, demonstrating its effectiveness in reducing redundant computation in diffusion models.
Limitations and availability
TeaCache is currently available for use on AMD Ryzen AI systems running Stable Diffusion 3 and 3.5. It requires the ONNX Runtime and is compatible with the Ryzen AI NPU's XDNA architecture. The method does not support all Stable Diffusion models, as it is tailored for transformer-based diffusion models like SD3 and SD3.5. The supplied announcement does not specify exact performance benchmarks or latency improvements, but it highlights the potential for reduced computation in diffusion models.
