skipper17/Analytic-DPM

Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models

Code for the paper Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models

News (Apr 22, 2022): Analytic-DPM received an Outstanding Paper Award at ICLR 2022!

Requirements

pytorch=1.9.0

Run experiments

You can change the phase variable in the code to determine the specific experiment you run.

For example, setting phase = "sample_analytic_ddpm" will run sampling using the Analytic-DDPM.

You can find all available phases in run_xxx.py.

CIFAR10

$ cd cifar_imagenet_codes
$ python run_cifar10.py

CelebA 64x64

$ cd celeba_lsun_codes
$ python run_celeba.py

Imagenet 64x64

$ cd cifar_imagenet_codes
$ python run_imagenet64.py

LSUN Bedroom

$ cd celeba_lsun_codes
$ python run_lsun_bedroom.py

Pretrained models and precalculated statistics

• CIFAR10 model: [checkpoint] trained by ourselves

• CelebA 64x64 model: [checkpoint] from https://github.com/ermongroup/ddim

• Imagenet 64x64 model: [checkpoint] from https://github.com/openai/improved-diffusion

• LSUN Bedroom model: [checkpoint] from https://github.com/pesser/pytorch_diffusion

• Precalculated Gamma vectors: [link]