Announcement_17

We present PICABench, a new benchmark and evaluation protocol for assessing physical realism in image editing — an often overlooked dimension in current generative models. PICABench systematically evaluates the physical consequences across eight sub-dimensions spanning optics, mechanics, and state transitions, with a reliable PICAEval protocol combining VLM-as-a-judge and region-level human annotations. We also build PICA-100K, a dataset for learning physics from videos. Evaluations show that physical realism remains a major challenge. PICABench aims to drive the next wave of physics-aware, causally consistent image editing. [Homepage] [GitHub] [ PICABench Dataset] [ PICA-100K Dataset] [Paper].