TerraVis world-consistency evaluation framework for text-to-image generation accepted at NeurIPS 2026

An AI-generated image can score high on aesthetics, prompt alignment, and human preference metrics while the picture hides distorted limbs, mutually penetrating objects, or unreasonable physical phenomena. As reported by Machine Heart, these errors are not always visible from existing scores, which exposes a blind spot in text-to-image evaluation: whether the objects, interactions, and spatial relationships in the picture conform to real-world constraints.

Accepted to the NeurIPS 2026 Evaluations and Datasets Track, TerraVis is precisely built around this problem, proposing a "world consistency" evaluation framework. It covers three levels and 18 categories of problems, and it not only gives a score but also points out the specific problem, visual evidence, and severity.

According to the definition in the report, world consistency focuses on whether the macro objects, interactions, and scenes in an image are reasonable at the structural, spatial, physical, biological, and common-sense levels. One key constraint is that the evaluation looks only at the image itself and does not depend on the generation prompt.

This clearly distinguishes it from "whether the prompt was drawn correctly." The report gives two cases pointing in opposite directions: an image of a computer desk received a TIFA alignment score of 1.00, but the laptop's structure was clearly distorted, and the human world-consistency rating was 0; another image deviated from the prompt requirement of "a fountain with no water flow" and drew a fountain spraying water, yet the picture itself conformed to reality, and both humans and TerraVis gave it the highest world-consistency rating.

A prompt may also actively request a scene that does not conform to reality, for example asking books in a library to float in the air. In this case, a lower world-consistency score does not mean the creation failed, but rather indicates that the image adopts a surreal setting. Alignment, preference, and world consistency need to be measured separately.

World consistency is also not equivalent to photographic realism. Cartoon simplified textures, flat colors in illustrations, and details omitted in sketches should not be treated as problems; but a hand that is holding a cup penetrating into the cup's body violates the spatial constraints between objects, regardless of art style. As long as an image contains recognizable macro objects or scenes, stylized images can also be evaluated; purely abstract patterns, icons, charts, and interface screenshots and other unsuitable content return N/A.