TeaCache speeds up Stable Diffusion 3.5 on AMD Ryzen AI

Researchers demonstrated that integrating TeaCache into Stable Diffusion 3 and 3.5 pipelines can accelerate image generation on AMD Ryzen AI neural processing units without retraining. By reusing intermediate transformer outputs across similar timesteps based on an accumulated relative L1 distance, the approach achieved average speedups ranging from 1.56x to 2.23x depending on the model and configuration. The authors mapped the technique to a static ONNX-based pipeline by running a separate norm module preemptively and handling caching decisions on the CPU. Open article →