Reproduce “NextGAN: Towards Simple and Scalable GAN Training for Image and Video Generation” 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

Xinhao Lou, Dongye Xie, Peng Sun, Chenyi Zhuang, Kaicheng Yu, Tao Lin

NEXTGAN trains a generator by supervising a critic’s gradient to predict real-minus-generated feature displacements at interpolated points, then updating the generator using the critic’s potential at generated samples. On text-to-video, a fine-tuned 1.3B Wan2.1 model scores 83.13 VBench Total with two sampling steps, above 1.3B methods, but not at four steps or across model scales.

Summary by alphaXiv.

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