Reproduce “Few-shot Image Generation via Cross-domain Correspondence” 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

Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A. Efros, Yong Jae Lee, Eli Shechtman, Richard Zhang

A method is proposed for few-shot image generation that adapts a pre-trained GAN to produce diverse, high-quality images from as few as 1-10 training samples. This approach prevents overfitting and mode collapse while also discovering cross-domain correspondences between source and target image domains.

Key insights

  • Preserving the structural relationships (diversity) learned from a large source dataset is crucial for preventing overfitting and mode collapse in few-shot GAN adaptation.
  • A dual discriminator strategy, combining an image-level discriminator for high-fidelity realism in specific regions and a patch-level discriminator for generalized local realism, enables diverse generation with limited data.
  • The proposed method can automatically discover one-to-one semantic correspondences between latent codes across different source and target domains, enabling consistent transformations.

Summary by alphaXiv.

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