dpm-solver

Why we included this project

Diffusion sampling is slow because it normally needs dozens of denoising steps, and DPM-Solver attacks exactly that bottleneck. This repo implements the DPM-Solver and DPM-Solver++ algorithms from the NeurIPS 2022 papers as high-order ODE solvers that bring the needed function evaluations down to around 10 to 20, with no retraining. Because it is model-agnostic and training-free, it slots in as a drop-in replacement sampler for discrete-time and continuous-time diffusion models alike; it also underlies the default multistep scheduler in Hugging Face Diffusers for Stable Diffusion pipelines. If you run guided-diffusion workloads such as text-to-image or inpainting on existing models and want them to feel snappier, this gives you a direct way to get that speed. The repo also includes working integration examples for Stable Diffusion, DeepFloyd-IF, and guided diffusion, so it works both as a solver to adopt and as a reference for wiring advanced samplers into real pipelines.