AI

TerraVis: Towards Evaluation of World-Grounded Visual Consistency in Text-to-Image Generation via MLLM Workflows

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/10/2026, 04:18:24 UTC.

Briefings

TerraVis is a new evaluation framework for text-to-image models that identifies world-grounded visual consistency violations, such as impossible anatomy or malformed structures. It uses an MLLM to assess model eligibility and scores images on 18 violation types, classifying them as minor or major. The framework achieved the strongest correlation with human judgments of world consistency among existing metrics across diverse models on two benchmarks.

Key points

• TerraVis defines a structured taxonomy of world-consistency violations spanning object-, interaction-, and scene-level failures.
• The framework achieves the strongest correlation with human judgments of world consistency among existing metrics across diverse open-source and proprietary models.

Context

Existing evaluation metrics fail to capture visually appealing images that violate real-world plausibility, such as impossible anatomy or physically implausible interactions. TerraVis addresses this gap by providing a systematic method to quantify and compare world-consistency failures across different models.

Limitations

The framework was tested on two widely used benchmarks with diverse open-source and proprietary models. Results are based on a single snapshot from October 2026, and the specific model names and benchmark datasets used are not detailed in the evidence.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 14/11/2026, 11:55:32 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeig2jvdmba54ugm2rrqkctiztu6uchozkn4limarbzjc3dlfszkady