AsyncDiff
AsyncDiff is a universal and plug-and-play acceleration framework for diffusion models that enables parallel processing across multiple devices through asynchronous denoising. Presented at NeurIPS 2024, this PyTorch-based solution breaks the sequential dependency in standard diffusion sampling by dividing the noise prediction workflow across available hardware. Users can achieve significant inference speedups, with benchmarks showing up to 2.8x faster generation on Stable Diffusion XL using four devices and 1.8x acceleration on AnimateDiff with two devices. The software supports a wide range of stability AI models and pipelines, including Stable Diffusion 3 Medium, version 2.1, and 1.5, as well as the SDXL base, Inpainting, and x4 Upscaler variants. Additionally, it integrates with ControlNet, Stable Video Diffusion, and AnimateDiff for text-to-image and image-to-video tasks. Designed for scalability, AsyncDiff allows researchers and developers to leverage multi-GPU setups to reduce generation time without co