DeepCache is a CVPR 2024 research project that accelerates diffusion models during inference by caching and reusing intermediate features across denoising steps. The method achieves up to 2.3x speedup on Stable Diffusion with minimal quality loss. It supports Stable Diffusion v1.5, SDXL, Stable Video Diffusion, and various pipelines including inpainting and img2img. The technique is training-free and integrates into the HuggingFace Diffusers library as a plug-and-play optimization.