WorldJen - Subject Consistency: leaderboard

WorldJen (2026) evaluates text-to-video generators on 50 adversarially curated prompts, with a Gemini 3 Flash judge answering 10 prompt-specific Likert questions per quality dimension on the full-resolution video. This board is the mean Likert score for whether the main subject keeps a consistent shape, colour and identity, on a 1 to 5 scale; higher is better.

Metric: Mean Likert score (1-5) for subject consistency (the main character or object keeps its shape, colour and identity), Gemini 3 Flash judge (gemini-3-flash-preview) viewing the video at native resolution and answering 10 prompt-specific Likert questions per dimension (1 major failure, 5 perfect execution) on the 50 curated WorldJen text-to-video prompts, dimensions a prompt does not exercise excluded; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Veo 3.1 Fast4.59
2Kling v2.6 Pro4.59
3Hunyuan v1.54.52
4Wan v2.2 A14B4.23
5LTX-24.12
6Wan 2.1 1.3B3.94

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