TerraVis: Towards Evaluation of World-Grounded Visual Consistency in Text-to-Image Generation via MLLM Workflows
Shuai Fu, Jing Gu, Jian Zhou, Zicheng Duan, Gengze Zhou, Qi Wu
Adelaide University · xAI · Responsible AI Research Centre
Published October 2, 2026 · on the feed October 9, 2026 · ▲ 4 · arXiv 2610.02959
The abstract
Recent text-to-image models have made substantial progress in photorealism, aesthetics, and text-image alignment. Yet visually appealing images can still violate real-world plausibility, exhibiting malformed object structures, impossible anatomy, physically implausible interactions, or inconsistent spatial relationships. Such failures are not well captured by existing fidelity, aesthetics, preference, or alignment metrics. To address this gap, we introduce TerraVis, a framework for evaluating world-grounded visual consistency in generated images. TerraVis defines a structured taxonomy of world-consistency violations spanning object-, interaction-, and scene-level failures, and employs a multi-stage evaluation framework to identify and quantify them. Given an image, TerraVis first uses an MLLM to assess its eligibility for evaluation, then detects violations across 18 taxonomy-defined types and classifies them as minor or major to derive an overall world-consistency score. Across diverse open-source and proprietary text-to-image models on two widely used benchmarks, TerraVis achieves the strongest correlation with human judgments of world consistency among existing metrics. Our benchmark results further show that models that achieve strong performance on conventional metrics can still exhibit substantial world-consistency failures. These findings highlight world consistency as a complementary evaluation dimension and demonstrate that TerraVis enables systematic quantification, diagnosis, and comparison of such failures. Our code is publicly available at https://github.com/ShyFoo/TerraVis.
The Brief
2:36A short podcast summary of the paper. Written and voiced by AI.
Transcript
TerraVis. Towards Evaluation of World-Grounded Visual Consistency in Text-to-Image Generation via MLLM Workflows. By Shuai Fu and five others, from Adelaide University, xAI, and one other institution. Published on the second of October. Its code repository is worked out: about seventy-five source files, with a readme and a licence.
Text-to-image models have advanced a long way in photorealism, in aesthetics, and in matching the image to the text. Yet an image can be beautiful and still break the rules of the real world. Objects come out with malformed structure. Anatomy is impossible. Interactions between things defy physics, or the spatial relationships in a scene do not agree with each other. The metrics in common use, whether they measure fidelity, aesthetics, preference or alignment, do not capture these failures well. A picture can score highly on all of them and still show a hand that no person could have.
TerraVis is a framework for evaluating what the authors call world-grounded visual consistency in generated images: whether what is depicted could plausibly exist in the real world. Its basis is a structured taxonomy of violations of world consistency. These span three levels. There are failures of single objects, failures of the interactions between them, and failures of the scene as a whole. The taxonomy gives the evaluation a fixed vocabulary for what can go wrong.
Evaluation runs in several stages. Given an image, TerraVis first uses a multimodal large language model to judge whether the image is eligible for evaluation at all. It then detects violations across eighteen types defined by the taxonomy. Each violation it finds is classified as either minor or major, and from those classifications comes an overall score for world consistency. The result is both a single number and a breakdown of what went wrong and how badly.
The framework is tested on a range of open-source and proprietary text-to-image models, using two widely used benchmarks. Among existing metrics, TerraVis has the strongest correlation with human judgments of world consistency. The benchmark results also show something about the models themselves. Those that do well on conventional metrics can still display substantial failures of world consistency. Scoring high on the familiar measures, in other words, does not guarantee an image that obeys the world.
The authors place world consistency as a complementary dimension of evaluation, alongside the existing ones rather than in place of them. With a taxonomy and a staged process behind it, TerraVis makes these failures something that can be quantified, diagnosed and compared across models, which turns a vague sense that an image looks wrong into a count of specific, classified errors.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free