Reproduce “Learning to Be Guided: Shaping Latent Geometry for Classifier-Free Guidance” 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

Haoming Liu, Junhao Zhu, Shaoyi Zheng, Yuanhe Guo, Jinnuo Liu, Shenji Wan, Hongyi Wen

The authors introduce Prototype Regularization (PR) to shape autoencoder latents for classifier-free guidance (CFG) in class-conditional image generation. PR compares pooled latent features with class prototypes, encouraging distinguishable class geometry before diffusion training. On ImageNet 256×256, with matched training budgets, PR reaches guided generation FID (gFID) 2.28, versus 3.31 for LDM; center-transplant controls support, but do not isolate, geometry’s contribution.

Summary by alphaXiv.

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