Reproduce “GeoComposer: Geometry-Grounded Photographic Composition Instruction” 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
Shuangzhi Li, Qiaoqiao Jia, Xingxin Chen, Guile Wu, Dongfeng Bai
GeoComposer gives advice and an example edit from a single image, using patch-level geometry supervision and rewards for geometry, image quality, and instruction following. Across 900 validation examples in three tasks, it achieves a 78.1% composition win rate versus Nano Banana Pro’s 74.9%, and MET3R of 0.184 versus 0.191. Each output is one exemplar; one view may omit context.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.