About this document
DeepCache: Training-Free Diffusion Acceleration by 398667135 is a document available to read on EtoBox.
The document introduces DeepCache, a novel approach to accelerate diffusion models without retraining, achieving speedups of 2.3× for Stable Diffusion v1.5 and 4.1× for LDM-4-G. By leveraging the temporal redundancy in high-level features during the denoising process, DeepCache reduces redundant computations, enhancing inference efficiency. The method demonstrates superior performance compared to traditional pruning and distillation techniques across various datasets and sampling methods.
- Author
- 398667135
- Language
- EN