Introducing DualStyleGAN, RQ‑VAE Transformer, and VFD: Recent CVPR 2022 Open‑Source Algorithms

Jack Cui introduces three recent CVPR 2022 papers that have been open‑sourced: DualStyleGAN, RQ‑VAE Transformer, and VFD.

DualStyleGAN

DualStyleGAN is a high‑resolution portrait style‑transfer GAN that can transfer the style of a reference image B onto a source image A, producing a stylized output C. The algorithm can be used to apply anime or other artistic styles to personal portraits.

It has gained over 300 stars on GitHub. The project is available at https://github.com/williamyang1991/DualStyleGAN . Users can try it via a web demo, a HuggingFace Space, a Colab notebook, or by setting up a local Conda environment. The Conda installation command is:

conda env create -f ./environment/dualstylegan_env.yaml

RQ‑VAE Transformer

RQ‑VAE Transformer combines a residual‑quantized VAE encoder with a transformer decoder to generate images from textual descriptions, achieving better results on text‑to‑image tasks. Example prompts include “A cheeseburger in front of a snow‑capped mountain range” and “a cherry blossom tree on the blue ocean”.

The method consists of two stages: the RQ‑VAE encodes images into discrete latent codes, and the RQ‑Transformer predicts those codes from text. The code is released at https://github.com/kakaobrain/rq-vae-transformer . Installation follows the requirements file, e.g.:

pip install -r requirements.txt

VFD

VFD is a deep‑fake detection tool that analyses images to determine authenticity, addressing the increasing realism of face‑swap technologies. The project, also from CVPR 2022, is open‑sourced at https://github.com/xaCheng1996/VFD .

These three tools illustrate recent advances in generative AI and forensic analysis, and they are readily accessible for experimentation.