High aesthetic score. High text alignment. High human preference. But look closer — the chair has three legs, the shadow falls the wrong way, the door handle floats in mid-air. Traditional metrics pick the "best" image, but humans judge it as physically wrong. There has been no benchmark for this gap — until now. TerraVis, accepted at NeurIPS 2026 Evaluations and Datasets Track, presents a "world consistency" evaluation framework (ArXiv 2610.02959). TerraVis evaluates generated images across three hierarchical levels and 18 problem categories, providing not just a score but specific issues, visual evidence, and severity ratings. The authors show that aesthetic, alignment, and preference metrics select different "best" images than human world-consistency judgments — high scores on existing metrics do not guarantee physical plausibility. An online demo lets users upload their own image and receive a world-consistency score with itemized reasoning. Code, dataset, and demo are all open. Result: TerraVis is the first world-consistency benchmark for text-to-image — 3 levels, 18 problem categories, score + visual evidence + severity — accepted at NeurIPS 2026, showing traditional metrics and human physical judgment diverge, with open code, dataset, and online demo. TerraVis: Towards Evaluation of World-Grounded Visual Consistency Paper: arxiv.org/abs/2610.02959 Code: github.com/ShyFoo/TerraVis Dataset: huggingface.co/datasets/ShyF… Demo: shyfoo.github.io/TerraVis-we… Our report: jiqizhixin.com/articles/2026… 📬 #PapersAccepted by Jiqizhixin