InnoAds: A Foundation Dataset and Benchmark for Advertising Visual Effects in Video Generation

Run Ling, Ke Cao, Runze He, Ao Ma, Xi Zhao, Jian Lu, Haowei Liu, Shuo Lu, Yun wang, Jing Wang, Yihua Shao, Zhanjie Zhang, Wei Feng, Zheng Zhang, Jingjing Lv, Junjie Shen, Ching Law, Guibing Guo

Keywords: Advertising Visual Effects, Video Generation, Dataset and Benchmark

TL;DR: We propose InnoAds-32K, a 32K image–text–video dataset, and InnoAds-Bench, a six-metric benchmark for advertising visual effects. Evaluations show SOTA video generation models fall short on Ad-VFX, setting a new research foundation.

Abstract:

Advertising visual effects (Ad-VFX) are the significant visual elements of advertising videos that combine dynamic product presentation with accompanying descriptive text.
However, research on Ad-VFX has been hindered by the lack of dedicated datasets and standardized evaluation protocols, as it is still an emerging domain.
To address this issue, we introduce InnoAds-32K, a foundation dataset of over 32,000 curated image–text–video triples tailored for advertising scenarios.
Furthermore, we propose InnoAds-Bench, a comprehensive benchmark that spans six evaluation dimensions: visual quality and text relevance as general metrics, and motion, product consistency, text stability, and creative rationality as advertising-specific metrics.
Based on this suite, we systematically evaluate state-of-the-art video generation models, revealing substantial limitations in their ability to satisfy advertising requirements.
In summary, InnoAds-32K and InnoAds-Bench provide the first standardized foundation for Ad-VFX video generation, paving the way for future research in advertising scenarios.

Primary Area: datasets and benchmarks

Submission Number: 3876