Embedded Characteristic Score for Evaluating Generative Image Models
The distributional match between images from a generative model and real images can be measured by embedding both sets with a pretrained neural network and comparing, feature by feature, the empirical characteristic functions of the two embedding distributions near the origin. Because the characteristic function near the origin encodes a distribution's moments and tails and stays stable even for heavy-tailed data, this score detects mismatches in higher moments and tails that Gaussian-based metrics like FID cannot.
2501.00744
Generative image models learn to produce samples from a target image distribution; evaluating them means measuring how well the generated distribution matches the real one. The dominant metric, Frech…