Reproduce “A Latent Distribution Perspective on Evaluating and Improving Latent Generative Models” 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
Xianghong Fang, Wenjie Shu, Tongda Xu, Wenlong Mou, Dehan Kong, Tim G. J. Rudner
The authors introduce generation-aware reconstruction (GAR) to test latent-image decoders between reconstruction and generation inputs. GAR perturbs an encoded image latent with noise, denoises it using the generative model, and measures output FID against source images. In mixed SiT-B/XL settings at noise level 0.8, GAR-FID correlates with gFID at 0.99/0.95, while rFID correlates at −0.30/−0.28. These empirical associations do not establish universal prediction; direct gFID remains necessary for evaluating final generation quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.