Reproduce “Making Reconstruction FID Predictive of Diffusion Generation FID” 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
Tongda Xu, Mingwei He, Shady Abu-Hussein, Jose Miguel Hernandez-Lobato, Chunhang Zheng, Kai Zhao, Chao Zhou, Ya-Qin Zhang and 1 more
Xu et al. from Tsinghua University and the University of Cambridge introduce Interpolated FID (iFID), a new metric to assess the quality of a Variational Autoencoder's latent space for latent diffusion model (LDM) generation. iFID achieves a strong positive correlation (PCC and SRCC approaching 0.9) with LDM generation FID (gFID) across diverse VAEs, effectively addressing the existing "reconstruction-generation dilemma" where traditional VAE metrics fail to predict LDM performance.
Key insights
- Standard reconstruction metrics (e.g., PSNR, SSIM, LPIPS, rFID) often exhibit weak or negative correlations with LDM generation quality (gFID), confirming the "reconstruction–generation dilemma."
- VAEs optimized solely for reconstruction tend to create disconnected latent spaces, which are detrimental for diffusion models that rely on smooth interpolation for generalization.
- Latent spaces that allow interpolations between nearby training samples to produce valid, realistic images are highly conducive to good diffusion generation, a property directly evaluated by iFID.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.