Reproduce “Looks Too Good To Be True: An Information-Theoretic Analysis of Hallucinations in Generative Restoration Models” on OpenResearch

About this paper

Regev Cohen, Idan Kligvasser, Ehud Rivlin, Daniel Freedman

Researchers from Verily AI introduce an information-theoretic framework to analyze hallucinations in generative image restoration models, demonstrating a fundamental tradeoff where improved perceptual quality inherently leads to increased uncertainty. This work provides a theoretical explanation for why models produce realistic details not present in the original data, showing that achieving perfect perception requires at least twice the inherent uncertainty of the inverse problem.

Key insights

• Achieving perfect perceptual quality in generative restoration requires at least twice the inherent uncertainty of the inverse problem, a lower bound that is independent of data distribution or model.
• The severity of the uncertainty-perception tradeoff intensifies in higher dimensions, meaning even small perceptual improvements in complex data like images come with a notable increase in uncertainty.
• The newly established uncertainty-perception tradeoff induces the well-known distortion-perception tradeoff, indicating that increased uncertainty as perception improves also corresponds to increased distortion.

Summary by alphaXiv.