TerraVis at NeurIPS 2026: New Benchmark Flags Physically Wrong AI Images That Score High
jiqizhixin · x · 2026-10-20
TerraVis (arXiv 2610.02959), accepted to the NeurIPS 2026 Evaluations and Datasets Track, introduces a "world consistency" evaluation framework for generated images: pictures with high aesthetic scores, strong text alignment, and human preference can still contain three-legged chairs, wrongly cast shadows, or floating door handles — and traditional metrics tend to select exactly these images.
- The framework evaluates across three hierarchical levels and 18 problem categories, outputting specific issues, visual evidence, and severity ratings rather than a single score.
- The authors show aesthetic, alignment, and preference metrics pick 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 images for evaluation.
More from Multimodal
- AI-Generated Music Video 'Mike from Jersey' Shows Cartoon Deli Scene — Ass_Blast69 · 2026-10-20
- Nutrient launches document decision model: answers in ~200ms with calibrated confidence — TheZachMueller · 2026-10-20
- New MULTIFINGROUND Benchmark Shows LLMs Recognize Financial Concepts but Fail at Full Evidence Extraction — TheFinAI · 2026-10-20
- MOSS model turns out remarkably good at inpainting, workflow shared on Civitai — Clasyc · 2026-10-20
- iPhone Action Button + Grok voice + HeyGen MCP turns a sentence into an avatar video — HeyGen · 2026-10-20
- AI-generated Tesla Roadster street video is indistinguishable from real footage — Tiny-Meet5202 · 2026-10-20