Reproduce “Contrastive Learning for Unpaired Image-to-Image Translation” 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
Taesung Park, Alexei A. Efros, Richard Zhang, Jun-Yan Zhu
This work introduces Contrastive Unpaired Translation (CUT), a method for unpaired image-to-image translation that replaces cycle-consistency with patch-based contrastive learning to preserve input content. CUT achieved lower FID scores and improved semantic correspondence compared to previous methods, while also demonstrating increased training efficiency.
Key insights
- Content preservation in unpaired image translation can be effectively achieved by maximizing mutual information between input and output patches using contrastive learning, removing the need for inverse mappings or cycle-consistency.
- Using internal negative samples (other patches from the same input image) is more effective for content preservation than external negatives, as it forces finer-grained distinctions.
- Applying contrastive loss across multiple layers of the encoder provides dense supervision, crucial for preserving both global structure and fine-grained details during image synthesis.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.