Project | Paper

PyTorch implementation of D2C: Diffuison-Decoding Models for Few-shot Conditional Generation.

Abhishek Sinha*, Jiaming Song*, Chenlin Meng, Stefano Ermon

Stanford University

Conditional generative models of high-dimensional images have many applications, but supervision signals from conditions to images can be expensive to acquire. This paper describes Diffusion-Decoding models with Contrastive representations (D2C), a paradigm for training unconditional variational autoencoders (VAEs) for few-shot conditional image generation. By learning from as few as 100 labeled examples, D2C can be used to generate images with a certain label or manipulate an existing image to contain a certain label. Compared with state-of-the-art StyleGAN2 methods, D2C is able to manipulate certain attributes efficiently while keeping the other details intact.

Here are some example for image manipulation. You can see more results here.

Attribute	Original	D2C	StyleGAN2	NVAE	DDIM
Blond
Red Lipstick
Beard

The code has been tested on PyTorch 1.9.1 (CUDA 10.2).

To use the checkpoints, download the checkpoints from this link, under the checkpoints/ directory.

# Requires gdown >= 4.2.0, install with pip
gdown https://drive.google.com/drive/u/1/folders/1DvApt-uO3uMRhFM3eIqPJH-HkiEZC1Ru -O ./ --folder

The main.py file provides some basic scripts to perform inference on the checkpoints.

We will release training code soon on a separate repo, as the GPU memory becomes a bottleneck if we train the model jointly.

Example to perform image manipulation:

• Red lipstick

python main.py ffhq_256 manipulation --d2c_path checkpoints/ffhq_256/model.ckpt --boundary_path checkpoints/ffhq_256/red_lipstick.ckpt --step 10 --image_dir images/red_lipstick --save_location results/red_lipstick

• Beard

python main.py ffhq_256 manipulation --d2c_path checkpoints/ffhq_256/model.ckpt --boundary_path checkpoints/ffhq_256/beard.ckpt --step 20 --image_dir images/beard --save_location results/beard

• Blond

python main.py ffhq_256 manipulation --d2c_path checkpoints/ffhq_256/model.ckpt --boundary_path checkpoints/ffhq_256/blond.ckpt --step -15 --image_dir images/blond --save_location results/blond

Example to perform unconditional image generation:

python main.py ffhq_256 sample_uncond --d2c_path checkpoints/ffhq_256/model.ckpt --skip 100 --save_location results/uncond_samples

We implement a D2C class here that contains an autoencoder and a diffusion latent model. See code structure here.

Useful functions include: image_to_latent, latent_to_image, sample_latent, manipulate_latent, postprocess_latent, which are also called in main.py.

• Release checkpoints and models for other datasets.
• Release code for conditional generation.
• Release training code and procedure to convert into inference model.
• Train on higher resolution images.

If you find this repository useful for your research, please cite our work.

@inproceedings{sinha2021d2c,
  title={D2C: Diffusion-Denoising Models for Few-shot Conditional Generation},
  author={Sinha*, Abhishek and Song*, Jiaming and Meng, Chenlin and Ermon, Stefano},
  year={2021},
  month={December},
  abbr={NeurIPS 2021},
  url={https://arxiv.org/abs/2106.06819},
  booktitle={Neural Information Processing Systems},
  html={https://d2c-model.github.io}
}

This implementation is based on:

• DDIM.
• NVAE.
• IDinvert