xDiT is a distributed inference system designed for Diffusion Transformers, enabling high-throughput image and video generation by leveraging parallelism strategies including PipeFusion, sequence parallel, and hybrid parallel methods. It addresses the quadratic computational complexity of attention mechanisms in DiTs by distributing workloads across multiple GPUs and machines to meet real-time inference requirements. The project integrates with ComfyUI and provides compilation and cache acceleration for single GPU execution.