TerraVis: Towards Evaluation of World-Grounded Visual Consistency in Text-to-Image Generation via MLLM Workflows
TerraVis uses multimodal LLMs to score world-consistency failures in text-to-image outputs across 18 violation types.
TerraVis evaluates whether text-to-image outputs obey real-world visual consistency rather than only aesthetics or prompt alignment. An MLLM workflow checks eligibility, detects 18 taxonomy-defined violations at object, interaction, and scene levels, and labels them minor or major to produce a world-consistency score. Across open-source and proprietary generators on two common benchmarks, TerraVis correlated best with human judgments among existing metrics. Models that score well on conventional metrics still showed substantial world-consistency failures.
- Taxonomy covers object, interaction, and scene-level failures
- MLLM workflow scores 18 violation types as minor or major
- Best correlation with human world-consistency judgments among compared metrics
- Strong conventional scores can still hide major plausibility failures
From the original article
1 min read at huggingface.co
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.
By Shuai Fu, Jing Gu, Jian Zhou · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.