Summary
Creating videos using AI often involves giving a description, or prompt, that guides what the video looks like and how it moves. The researchers point out that understanding how well these original prompts can be recovered from the videos themselves is important but hasn’t been measured before. They built a large test called VI-Bench that checks if AI models can figure out the exact prompts used by looking at videos generated from them. Their tests show that even the best AI models have a hard time perfectly recovering these prompts, especially when the videos involve complex styles or multiple scenes. This means that figuring out the original instructions behind AI-generated videos is a tricky problem that needs more work.
AIGCprompt inversionvideo generationbenchmarksemantic groundingstyle controlcamera behaviorvideo language modelsmulti-shot compositionality
Authors
Wulin Xie, Rui Zhao, Kecen Li, Xiujin Liu, Bokang Zhang, Zheng Liu, Xinwen Hou, Chen Gong
Abstract
Recent advances in video generation have made prompt-based control increasingly central to AIGC video generation. Prompts specify what a video should depict and how it should be represented, controlling factors such as visual style or camera behavior. Understanding this recoverability is important both for creative reuse and editing, and for assessing prompt leakage risks. However, existing video understanding benchmarks do not measure this capability: a caption may describe what is visible, but a replayable prompt must recover the generation-relevant controls needed to reproduce the video. To address this gap, we introduce VI-Bench, a benchmark built from 16.1 million real-user prompts and 900 human-verified AIGC videos. VI-Bench spans three progressively harder settings, namely single-shot semantic grounding, control over style and camera behavior, and multi-shot compositional inversion, and evaluates five generation-critical dimensions: subject, action, scene, style, and camera. We evaluate 18 representative VLMs, including 2 proprietary and 16 open-source models on VI-Bench, using an Inversion Score that measures prompt-level alignment with the original prompt and video-level fidelity of the regenerated video. The results reveal substantial limitations: even the strongest model achieves only 0.632 on Inversion Score, performance degrades sharply as samples require richer control and multi-shot reasoning, and models often produce plausible prompts whose regenerated videos deviate from the reference. These findings show that video prompt inversion is a distinct and under-evaluated capability requiring models to transform visual understanding into replay-stable generative control.