AnimationBench - Semantic Consistency: leaderboard

Agreement of generated animations with the prompt and conditioning image in objects, actions, colors and scene, averaged over the four, on AnimationBench, an animation image-to-video benchmark (2026) built on IP preservation and the Twelve Basic Principles of Animation, about 30 generated videos per dimension, scored by Qwen3-VL-Max yes/no question sets. Percent; higher is better.

Metric: Semantic Consistency overall: mean of object, action, color and scene agreement with the prompt and conditioning image (the paper's Sem. Overall) (%), animation-style image-to-video generation from the AnimationBench close-set suite (fixed source images and prompts, about 30 videos per dimension per model), scored by Qwen3-VL-Max yes/no question sets aggregated per video; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1Veo 3.189.31
2Seedance 2.089.05
3Kling 2.686.11
4Seedance 1.0 Pro69.86
5Wan2.2-I2V-A14B67.86
6FramePack33.33
7HunyuanVideo-1.524.2

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