Reproduce “Return of Unconditional Generation: A Self-supervised Representation Generation Method” 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
Tianhong Li, Dina Katabi, Kaiming He
MIT CSAIL researchers propose Representation-Conditioned Generation (RCG), a framework that leverages self-supervised representations to achieve state-of-the-art unconditional image generation quality. This method rivals class-conditional models without using human labels, reaching an FID of 2.15 on ImageNet 256x256, and generalizes across various generative architectures.
Key insights
- Decomposing unconditional image generation into generating self-supervised representations and then conditioning image synthesis on them effectively bridges the semantic gap without human labels.
- Self-supervised representations provide sufficient semantic information to guide high-quality image generation, even outperforming discrete class labels in some cases.
- The modular RCG framework allows for flexible integration with various existing image generation architectures and enables efficient transfer learning for conditional generation tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.