Official implementation of the paper "Distilling ODE Solvers of Diffusion Models into Smaller Steps" at CVPR 2024
Our codebase consists of two parts depending on diffusion models.
The noise prediction models with DDIM, iPNDM, DPM-Solver, and DEIS is based on the code repositories of DPM-solver and DEIS. Check the code at dpm_solver.
The data prediction models with DDIM and EDM is heavily based on EDM codebase. Check the code at edm.
Please follow their instructions to set up each environment and download pretrained models. In launch.sh of each part, you will find a command to run each sampler on Cifar-10.
In noise prediction models, you need jax library to run DEIS. Please run follow command after you set up your environment referring to DPM-Solver repository.
# for pytorch user
pip install "jax[cpu]"