Repository files navigation

Jonathan Ho, Ajay Jain, Pieter Abbeel

Paper:https://arxiv.org/abs/2006.11239

Website:https://hojonathanho.github.io/diffusion

Experiments run on Google Cloud TPU v3-8.Requires TensorFlow 1.15 and Python 3.5, and these dependencies for CPU instances (seerequirements.txt):

pip3 install firepip3 install scipypip3 install pillowpip3 install tensorflow-probability==0.8pip3 install tensorflow-gan==0.0.0.dev0pip3 install tensorflow-datasets==2.1.0

The training and evaluation scripts are in thescripts/ subdirectory.The commands to run training and evaluation are in comments at the top of the scripts.Data is stored in GCS buckets. The scripts are written to assume that the bucket names are of the formgs://mybucketprefix-us-central1; i.e. some prefix followed by the region.The prefix should be passed into the scripts using the--bucket_name_prefix flag.

Models and samples can be found at:https://www.dropbox.com/sh/pm6tn31da21yrx4/AABWKZnBzIROmDjGxpB6vn6Ja

If you find our work relevant to your research, please cite:

@article{ho2020denoising,    title={Denoising Diffusion Probabilistic Models},    author={Jonathan Ho and Ajay Jain and Pieter Abbeel},    year={2020},    journal={arXiv preprint arxiv:2006.11239}}