Reproduce “Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now” 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
Ayush Sarkar, Hanlin Mai, Amitabh Mahapatra, Svetlana Lazebnik, D.A. Forsyth, Anand Bhattad
Generative models, despite producing visually convincing outputs, demonstrate systematic discrepancies in adhering to fundamental principles of projective geometry and illumination physics. Researchers at the University of Illinois Urbana-Champaign and Toyota Technological Institute at Chicago developed geometry-based classifiers that accurately detect these inconsistencies in "hard" generated images, achieving AUCs up to 0.84 on images that bypass pixel-based detectors.
Key insights
- Current generative models consistently fail to accurately reproduce fundamental projective geometry and illumination physics over long spatial scales within generated images.
- Geometric inconsistencies serve as reliable, model-agnostic artifacts for detecting synthetic content, even when pixel-level features are indistinguishable from real images.
- Addressing these systematic geometric flaws will likely require "structural innovation" in generative model architectures, moving beyond simply increasing data or model capacity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.