Official implementation of Diffusion Autoencoders

A CVPR 2022 paper:

Preechakul, Konpat, Nattanat Chatthee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. 2021. “Diffusion Autoencoders: Toward a Meaningful and Decodable Representation.” arXiv [cs.CV]. arXiv. http://arxiv.org/abs/2111.15640.

Usage

Note: Since we expect a lot of changes on the codebase, please fork the repo before using.

Quick start

A jupyter notebook.

For unconditional generation: sample.ipynb

For manipulation: manipulate.ipynb

Checkpoints

We provide checkpoints for the following models:

1. DDIM: FFHQ128 (72M, 130M), Bedroom128, Horse128
2. DiffAE (autoencoding only): FFHQ256, FFHQ128 (72M, 130M), Bedroom128, Horse128
3. DiffAE (with latent DPM, can sample): FFHQ256, FFHQ128, Bedroom128, Horse128
4. DiffAE's classifiers (for manipulation): FFHQ256's latent on CelebAHQ, FFHQ128's latent on CelebAHQ

Checkpoints ought to be put into a separate directory checkpoints. Download the checkpoints and put them into checkpoints directory. It should look like this:

checkpoints/
- bedroom128_autoenc
    - last.ckpt # diffae checkpoint
    - latent.ckpt # predicted z_sem on the dataset
- bedroom128_autoenc_latent
    - last.ckpt # diffae + latent DPM checkpoint
- bedroom128_ddpm
- ...

LMDB Datasets

We do not own any of the following datasets. We provide the LMDB ready-to-use dataset for the sake of convenience.

• FFHQ
• CelebAHQ
• LSUN Bedroom
• LSUN Horse

The directory tree should be:

datasets/
- bedroom256.lmdb
- celebahq256.lmdb
- ffhq256.lmdb
- horse256.lmdb

You can also download from the original sources, and use our provided codes to package them as LMDB files. Original sources for each dataset is as follows:

• FFHQ (https://github.com/NVlabs/ffhq-dataset)
• CelebAHQ (https://github.com/switchablenorms/CelebAMask-HQ)
• LSUN (https://github.com/fyu/lsun)

The conversion codes are provided as:

data_resize_bedroom.py
data_resize_celebhq.py
data_resize_ffhq.py
data_resize_horse.py

Google drive: https://drive.google.com/drive/folders/1abNP4QKGbNnymjn8607BF0cwxX2L23jh?usp=sharing

Training

We provide scripts for training & evaluate DDIM and DiffAE (including latent DPM) on the following datasets: FFHQ128, FFHQ256, Bedroom128, Horse128, Celeba64 (D2C's crop). Usually, the evaluation results (FID's) will be available in eval directory.

Note: Most experiment requires at least 4x V100s during training the DPM models while requiring 1x 2080Ti during training the accompanying latent DPM.

FFHQ128

# diffae
python run_ffhq128.py
# ddim
python run_ffhq128_ddim.py

FFHQ256

We only trained the DiffAE due to high computation cost. This requires 8x V100s.

sbatch run_ffhq256.py

After the task is done, you need to train the latent DPM (requiring only 1x 2080Ti)

python run_ffhq256_latent.py

Bedroom128

# diffae
python run_bedroom128.py
# ddim
python run_bedroom128_ddim.py

Horse128

# diffae
python run_horse128.py
# ddim
python run_horse128_ddim.py

Celeba64

This experiment can be run on 2080Ti's.

# diffae
python run_celeba64.py