AdSpark: A Large-Scale Dataset and Benchmark for Product-Centric Advertisement Video Generation

AdSpark. A Large-Scale Dataset and Benchmark for Product-Centric Advertisement Video Generation. By Zhifei Yang and eight others, from Peking University and JD dot com. Published on the seventh of October.

What now exists is AdSpark, a large-scale dataset and benchmark for product-centric advertisement video generation. The task asks for two things at once. A promotional video has to preserve the product's fine-grained identity, so the item on screen is recognisably the item being sold. And it has to present the product's selling points through a coherent narrative that runs across several shots. It is an emerging task, and the authors say it has been underexplored for a plain reason: there were no large advertisement-specific datasets to learn from, and no comprehensive framework for judging the results.

The dataset, AdSpark-300K, holds approximately three hundred thousand triplets, each made of a reference image, a prompt, and a video. It has a real-world subset and a synthetic subset. Every sample carries structured advertisement annotations: product identity, descriptions of selling points, creative plans, and aligned audio scripts. Those annotations line up with the two halves of the task. Identity annotations serve product preservation, while selling points, creative plans and scripts serve advertisement-oriented visual storytelling.

The data is based on material from a major e-commerce platform, so the products and advertisements come from real commercial use. Alongside the dataset sits AdSpark-Bench, a diagnostic benchmark that scores generated advertisements on six dimensions: visual quality, product fidelity, instruction adherence, temporal coherence, audio alignment, and advertisement effectiveness. The list runs from general video quality through to whether the result works as an advertisement at all.

The authors offer it as a unified dataset and benchmark for future research. Evaluating representative models on AdSpark-Bench exposes where current systems struggle: preserving the product, telling a story across multiple shots, and visualising selling points. Those are the same demands the annotations were built around. Models fine-tuned on AdSpark-300K are used to validate that the dataset is effective.

The pairing is the contribution: a dataset large and richly annotated enough to train on, and a benchmark that breaks a commercial video into the qualities an advertiser cares about, so that a model's shortfalls can be named rather than merely sensed. One note on availability: the authors say the dataset will be released upon acceptance.