Reproduce “Your GAN is Secretly an Energy-based Model and You Should use Discriminator Driven Latent Sampling” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Tong Che, Ruixiang Zhang, Jascha Sohl-Dickstein, Hugo Larochelle, Liam Paull, Yuan Cao, Yoshua Bengio

The authors introduce discriminator-driven latent sampling (DDLS), which uses discriminator gradients and Langevin dynamics to guide latent codes before image generation. It moves a generator’s inputs rather than rejecting outputs. On CIFAR-10, DDLS raises an unconditional pretrained SN-GAN’s Inception Score from 8.22 ± 0.05 to 9.09 ± 0.10 and lowers FID from 21.7 to 15.42, a result specific to this model and setup.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.