AdSpark: A Large-Scale Dataset and Benchmark for Product-Centric Advertisement Video Generation
AdSpark packages roughly 300,000 product image, prompt, and video triplets for training ad-focused video models.
The paper frames product-centric ad generation as a task where models must preserve product identity while building coherent multi-shot promotional stories. Its dataset includes real-world and synthetic subsets, with annotations for product identity, selling points, creative plans, and aligned audio scripts. The authors also introduce AdSpark-Bench, which scores outputs on visual quality, product fidelity, instruction following, temporal coherence, audio alignment, and ad effectiveness. They say benchmark results expose remaining weaknesses in product preservation, multi-shot storytelling, and selling-point visualization. HF Daily Papers' note
The paper frames product-centric ad generation as a task where models must preserve product identity while building coherent multi-shot promotional stories. Its dataset includes real-world and synthetic subsets, with annotations for product identity, selling points, creative plans, and aligned audio scripts. The authors also introduce AdSpark-Bench, which scores outputs on visual quality, product fidelity, instruction following, temporal coherence, audio alignment, and ad effectiveness. They say benchmark results expose remaining weaknesses in product preservation, multi-shot storytelling, and selling-point visualization. HF Daily Papers' note
score 5