DistriFusion is an algorithm that requires no training and takes advantage of multiple GPUs to accelerate diffusion model inference without compromising image quality. DistriFusion can reduce delay based on the number of devices used while maintaining visual fidelity.