DPM-Solver for DDPM and Guided-Diffusion

This is the example code for DPM-Solver applying to DDPM and Guided-Diffusion. The code is heavily based on DDIM codebase.

Dataset, Checkpoint and FID Stats

We support various datasets and checkpoints. Please check the config files in configs/.

Some checkpoints will be automatically downloaded in ~/ddpm_ckpt/, please check this code for details: functions/ckpt_util.py. Other checkpoints needs to be put in the model.ckpt_dir in the config file.

Config File	Checkpoint	FID Stats
cifar10.yml	Automatically download DDPM checkpoint on CIFAR-10 in ~/ddpm_ckpt/diffusion_models_converted/.	Download in ./fid_stats/fid_stats_cifar10_train_pytorch.npz
celeba.yml	Please download DDIM checkpoint on Celeb-A in ~/ddpm_ckpt/celeba/ckpt.pth.	Download in ./fid_stats/fid_stats_celeba64_train_50000_ddim.npz
imagenet64.yml	Please download improved-DDPM checkpoint on unconditional ImageNet64 in ~/ddpm_ckpt/imagenet64/imagenet64_uncond_100M_1500K.pt.	Download in ./fid_stats/fid_stats_imagenet64_train.npz
bedroom_guided.yml	Please download guided-diffusion checkpoint on unconditional LSUN bedroom in ~/ddpm_ckpt/bedroom/lsun_bedroom.pt.	Download in ./fid_stats/VIRTUAL_lsun_bedroom256.npz
imagenet128_guided.yml	Please download guided-diffusion diffusion checkpoint in ~/ddpm_ckpt/imagenet128/128x128_diffusion.pt, and guided-diffusion classifier checkpoint in ~/ddpm_ckpt/imagenet128/128x128_classifier.pt.	Download in ./fid_stats/VIRTUAL_imagenet128_labeled.npz
Imagenet256_guided.yml	Please download guided-diffusion diffusion checkpoint in ~/ddpm_ckpt/imagenet256/256x256_diffusion.pt, and guided-diffusion classifier checkpoint in ~/ddpm_ckpt/imagenet256/256x256_classifier.pt.	Download in ./fid_stats/VIRTUAL_imagenet256_labeled.npz
Imagenet512_guided.yml	Please download guided-diffusion diffusion checkpoint in ~/ddpm_ckpt/imagenet512/512x512_diffusion.pt, and guided-diffusion classifier checkpoint in ~/ddpm_ckpt/imagenet512/512x512_classifier.pt.	Download in ./fid_stats/VIRTUAL_imagenet512.npz

Example Script

We provide three example scirpt in sample.sh, including:

• Unconditional sampling on CIFAR-10 with linear beta schedule.

• Unconditional sampling on ImageNet64 with cosine beta schedule.

• Classifier-guided sampling on ImageNet256 with guidance scale 8.0 and dynamic thresholding.