XTC-Bench: leaderboard

XTC-Bench (2026) checks whether a unified multimodal model's image generation and visual understanding agree: from 2,000 COCO and Visual Genome scene graphs it derives a drawing prompt and per-fact questions, then compares the model's generated facts with its own answers. Accuracy-weighted cross-task agreement (0-1, shown times 100); higher is better.

Metric: Accuracy-weighted continuous cross-task agreement (AW-CCTA, 0-1, times 100): per-fact agreement between the model's generated image (scene graph extracted and judged) and its own answers to the matching questions, weighted by their joint correctness, averaged over all facts of XTC-Bench's 2,000 scene-graph-annotated images from COCO 2017 val and Visual Genome (over 31,000 atomic facts about objects, attributes and relations); a Qwen3-235B judge scores each fact 0-5, normalized to 0-1 and shown times 100; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Flash Image62.3
2BAGEL-7B-MoT56.9#4619
3BLIP3-o8B42.3
4Janus-Pro-7B39#6857
5OmniGen237.7#4579
6Tar-7B35.8
7Show-o2-7B33.3
8MMaDA-8B14.4
9Show-o5.4

Interactive version: theaggregate.ai/benchmark?slug=xtc-bench · How It Works · Data refreshed daily, snapshot 2026-11-21.