Reproduce “Localizing Memorized Regions in Diffusion Models via Coordinate-Wise Curvature Differences” 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
Gwangho Kim, Sungyoon Lee
The authors map curvature coordinate-wise to locate partial copying in text-to-image diffusion outputs. They estimate curvature from score derivatives and subtract an unconditional or less-trained reference to reduce ordinary structural signals. On SD v1.4 “All,” unconditional-baseline curvature difference reaches IoU 0.953, versus 0.564 for Bright Ending; this best threshold-sweep result is not universal: Bright Ending leads on SD v2.1 “TV + Non-mem.”
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.