COMEX: A Composition-Grounded Benchmark and Learning Framework for Explainable Aesthetic Image Cropping
The episode discusses COMEX, a paper presenting a benchmark and learning framework for explainable aesthetic image cropping. The authors propose outputting a crop box, composition category, and explanation grounded in photographic principles. They use an IO-reversal trick to create a large dataset and employ an SFT plus GRPO training method that significantly improves performance on real photos.