distrifuser

Why we included this project

DistriFusion tackles a specific pain point: high-resolution diffusion models like SDXL are slow on a single GPU, and the usual fix, retraining, is a big lift. Instead it spreads the work across several GPUs by splitting the image into patches, one per device, and reusing activations from the previous denoising step to hide the communication cost. That yields near-linear speedups without a visible quality hit, and the gains grow as resolution climbs. The code mirrors the Hugging Face diffusers API, so you swap in a DistriSDXLPipeline, set a DistriConfig, and launch with torchrun rather than rewriting your generation pipeline. It was a CVPR 2024 highlight, and NVIDIA folding the approach into TensorRT-LLM, with ColossalAI also supporting it, suggests the technique holds up beyond this reference implementation. If you already run multi-GPU PyTorch, this is about the least disruptive way to speed up diffusion inference.