VI-Bench (Video Prompt Inversion): leaderboard
VI-Bench (2026) tests video prompt inversion: a vision-language model watches an AI-generated video and writes a prompt that should regenerate it with the original video generator. The Inversion Score averages a GPT-4o judgment of the recovered prompt against the original prompt with an agent judgment of the regenerated video against the reference, each on a 0 to 1 scale, over 900 videos; higher is better.
Metric: Inversion Score (0-1; mean of the Prompt Score, GPT-4o (gpt-4o-2024-11-20) judging the recovered prompt against the original one on subject, action, scene, style and camera, and the Video Score, a Qwen3-VL-8B memory agent and Qwen3.5-9B judge agent comparing the video regenerated from the recovered prompt by the original generator (Wan2.2 or HunyuanVideo 1.5, fixed seed) with the reference, each a 1-5 rating normalized to 0-1; all 900 human-verified AIGC videos). Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Seed 2.0 Pro | 0.63 | #209 |
| 2 | GPT-4o (2024-11-20) | 0.58 | #864 |
| 3 | Qwen 2.5 VL 7B | 0.52 | #1796 |
| 4 | Qwen 3.5 9B | 0.52 | #803 |
| 5 | InternVL3-8B | 0.39 | #1695 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vi-bench-video-prompt-inversion · How It Works · Data refreshed daily, snapshot 2026-11-12.