Jiaming Song, Chenlin Meng and Stefano Ermon, Stanford

Implements sampling from an implicit model that is trained with the same procedure as Denoising Diffusion Probabilistic Model, but costs much less time and compute if you want to sample from it (click image below for a video demo):

The code has been tested on PyTorch 1.6.

Training is exactly the same as DDPM with the following:

python main.py --config {DATASET}.yml --exp {PROJECT_PATH} --doc {MODEL_NAME} --ni

python main.py --config {DATASET}.yml --exp {PROJECT_PATH} --doc {MODEL_NAME} --sample --fid --timesteps {STEPS} --eta {ETA} --ni

where

• ETA controls the scale of the variance (0 is DDIM, and 1 is one type of DDPM).
• STEPS controls how many timesteps used in the process.
• MODEL_NAME finds the pre-trained checkpoint according to its inferred path.

If you want to use the DDPM pretrained model:

python main.py --config {DATASET}.yml --exp {PROJECT_PATH} --use_pretrained --sample --fid --timesteps {STEPS} --eta {ETA} --ni

the --use_pretrained option will automatically load the model according to the dataset.

We provide a CelebA 64x64 model here, and use the DDPM version for CIFAR10 and LSUN.

If you want to use the version with the larger variance in DDPM: use the --sample_type ddpm_noisy option.

Use --interpolation option instead of --fid.

Use --sequence option instead.

The above two cases contain some hard-coded lines specific to producing the image, so modify them according to your needs.

@article{song2020denoising,
  title={Denoising Diffusion Implicit Models},
  author={Song, Jiaming and Meng, Chenlin and Ermon, Stefano},
  journal={arXiv:2010.02502},
  year={2020},
  month={October},
  abbr={Preprint},
  url={https://arxiv.org/abs/2010.02502}
}

This implementation is based on / inspired by:

• https://github.com/hojonathanho/diffusion (the DDPM TensorFlow repo),
• https://github.com/pesser/pytorch_diffusion (PyTorch helper that loads the DDPM model), and
• https://github.com/ermongroup/ncsnv2 (code structure).