TerraVis: a NeurIPS 2026 framework scores text-to-image world consistency across 18 failure types

机器之心 · wechat · 2026-10-12

Researchers from the University of Adelaide's AIML and xAI present TerraVis (NeurIPS 2026 Datasets Track), tackling a blind spot in text-to-image evaluation: images can score high on aesthetics, alignment, and preference metrics while containing twisted limbs, penetrating objects, or impossible physics.

Key points:

World-grounded visual consistency judges the image itself, independent of the prompt — an image violating "fountain with no water flow" can still score perfectly on world consistency
18 check types across three levels: object (structure/anatomy/text), interaction (contact/support/motion/optics), and scene (relative scale/occlusion/continuity), with a four-step MLLM workflow turning vague violations into concrete visual questions
On COCO-T2I and GenAI-Bench with GPT-5.5, TerraVis reaches Spearman correlations of 0.44/0.43 with human judgments (best baseline: 0.29/0.37); 48.6%/69.7% of sampled images had world-consistency issues per human majority vote
Rankings shift: GPT Image 1.5 leads traditional metrics, but both TerraVis and humans rank FLUX.2-dev first

An online demo lets users upload images and inspect per-item violations with visual evidence.

Original post →

TerraVis: a NeurIPS 2026 framework scores text-to-image world consistency across 18 failure types

More from Multimodal