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

AI summary

Overview

Research area: Computer vision and generative video modeling, specifically reference-conditioned video generation for e-commerce advertising.

Technical level: Advanced. The paper assumes familiarity with diffusion-based video generation, reference-to-video conditioning, LoRA fine-tuning, and automated video/audio evaluation metrics.

Scope: The paper introduces AdSpark-300K, a 300K-sample annotated product advertisement video dataset, and AdSpark-Bench, a six-dimension diagnostic benchmark for product-centric advertisement video generation (PC-AVG), together with an evaluation of 20 video generation models.

What This Paper Is About

Product-centric advertisement video generation requires a model to take a product reference image and a written ad plan, then produce a promotional video that keeps the product's fine-grained identity intact while telling a coherent multi-shot story and actually showing the selling points. Existing video datasets and benchmarks are built for general text-to-video, subject-to-video, or image-to-video tasks and lack advertising-specific annotations, so this paper builds a dedicated dataset with structured ad annotations and a benchmark that judges generated videos as advertisements, not just as videos.

Key Contributions

1. AdSpark-300K, described as the first large-scale dataset for PC-AVG: approximately 300K reference image–prompt–video triplets, split into a 100K real-world subset and a 200K synthetic subset, covering 40 product categories, over 3,000 sub-categories, and 651 hours of video at 720P, with structured annotations for product identity, selling points, creative plans (scene, style, shot design), and aligned audio scripts.

2. AdSpark-Bench, a diagnostic benchmark of 220 product-conditioned test cases that are SKU-disjoint from the training set, extending conventional video evaluation with advertisement-specific criteria including product fidelity, shot compliance, selling-point realization, and advertisement effectiveness.

3. Comprehensive model evaluation of proprietary and open-source video generation models on AdSpark-Bench, revealing limitations in product preservation, multi-shot storytelling, and selling-point visualization.

4. A fine-tuning validation study, where LTX-2 is adapted on AdSpark-300K with LoRA (rank 64) to produce LTX-AdSpark, demonstrating that the dataset measurably improves advertisement-oriented generation.

Main Findings

• Fine-tuning on AdSpark-300K substantially improves advertisement capabilities. Relative to base LTX-2, LTX-AdSpark raises cross-shot product consistency from 33.30 to 93.59, shot structure alignment from 46.29 to 77.53, and narrative coherence from 33.77 to 89.89.

• Closed-source models lead overall but still fail at product fidelity under motion. The paper reports that motion blur, shape distortion, and appearance drift degrade subject consistency, key-region similarity, and text fidelity even for stronger models.

• Open-source models show more pronounced weaknesses in instruction adherence and advertisement effectiveness, and often struggle with coherent multi-shot generation, reflected in weaker cross-shot product consistency and temporal coherence.

• The dataset is heavily multi-shot in structure. Over 85% of AdSpark-300K samples contain multiple shots, predominantly 2-shot (47.3%) and 3-shot (23.6%) compositions.

• The benchmark is balanced across shot counts. AdSpark-Bench contains 45 one-shot, 131 two-shot, and 44 three-shot cases, yielding 175 multi-shot cases for cross-shot consistency, transition quality, and narrative coherence evaluation.

• Different models specialize in different dimensions. For example, Seedance 2.0 reaches 34.31 on Text Fidelity and 94.99 on cross-shot consistency, HappyHorse-1.1 reaches 95.43 on Transition Naturalness and 89.41 on Shot Structure Alignment, and MV-S2V reaches the reported top scores of 43.44 on Aesthetic Score and 71.57 on Subject Consistency. LTX-AdSpark records 51.92 on Background Sound Consistency and 84.41 on Narration Script Consistency.

• Audio behavior was a clear dividing line. Several models (ViduQ2, Pixverse V5, VACE, SkyReels-V3, VINO, Refalign, Kaleido, HunyuanCustom, Bernini, MV-S2V, MAGREF) show dashes in the audio alignment columns, which the paper states indicate unavailable or invalid audio outputs excluded from evaluation.

• Ablations on real-world versus synthetic composition and training scale were run, with results reported in Tab. 6; the specific numbers are not included in the provided text.

Methodology in Plain English

Building the real-world subset. The team pulled product assets from a major e-commerce platform's product detail pages using authorized internal interfaces, starting from roughly 600K top-ranked SKUs chosen from platform-wide click-through statistics dated April 15, 2026, keeping videos between 10 and 1,000 seconds at 24 or 25 fps. That filtering left 192,077 SKUs with at least one eligible ad video, totaling 218,837 video entries, of which 26,760 SKUs had multiple eligible videos. They used SAM2.1-Hiera-Large to track each product through its video starting from the alpha channel of a transparent-background reference image, then threw out frames where the product mask was too small (under 5% of the frame), too large (over 80%), touching the frame boundary within a 2% margin, or unstable (adjacent-frame mask-area ratio above 3.0, or normalized centroid displacement above 0.35). Surviving consecutive frames were grouped into 5–20 second candidate clips and passed to Qwen3-VL-235B for identity, visibility, and overlay-artifact checks. Accepted clips were split into shots with TransNetV2; Qwen3-VL-235B wrote product identity, selling point, and creative plan annotations; Whisper-Large-v3-Turbo transcribed audio and aligned narration to shots. Twenty trained annotators then manually verified or rejected each sample.

Building the synthetic subset. Because good real advertisements are scarce, they also generated ads. Starting again from 600,000 click-ranked candidate SKUs, quality filtering required a transparent-background reference image of at least 800×800 resolution with a valid transparency channel, removing 199,870 SKUs and keeping 400,130. Suitability filtering used Qwen3-VL-235B to require high generation-suitability with no more than four primary subjects and four visible items and no overlays, watermarks, humans, or stray scene elements, removing 49,290 SKUs and keeping 350,840. Category filtering then excluded non-physical, service, medical, second-hand, and other risky categories. For retained products, GPT-5.5 wrote structured ad annotations that translate selling points into visualizable actions, with Gemini-3.1-Pro-Preview acting as an independent reviewer in an iterative review-and-refinement loop. Videos were then synthesized with multiple models — Seedance 2.0, Happy Horse, and Kling — to improve diversity and reduce model-specific bias, followed by human inspection for identity drift, artifacts, weak selling-point presentation, and audio-visual mismatch.

Evaluating. AdSpark-Bench segments each generated video into shots with TransNetV2 and greedily matches detected shots to planned shots by temporal IoU, sharing those correspondences across shot-structure, shot-execution, and cross-shot evaluations. Visual quality uses Aesthetic Score and MUSIQ. Product fidelity uses GroundingDINO and SAM2 to isolate foreground regions at several granularities, then DINOv3 feature similarity, RapidOCR text matching with minimum edit distance, and cross-shot consistency across cuts. Instruction adherence combines GPT-5.5 judgments of shot type/motion (with optical flow cues), shot count and boundary accuracy, scene/style/selling-point realization, and a GmeScore from gme-Qwen2-VL-7B-Instruct. Temporal coherence combines motion amplitude, motion smoothness (coefficient of variation of optical flow magnitudes), product motion stability, GPT-5.5 transition naturalness, and SSIM-based boundary checks. Audio alignment uses LAION-CLAP for background sound and Whisper-Large-v3-Turbo with character error rate for narration scripts. Advertisement effectiveness is judged by GPT-5.5 role-playing a potential customer on attractiveness, creative quality, and narrative coherence.

Models compared. ViduQ2, ViduQ3, Seedance 2.0, HappyHorse-1.1, Pixverse V5, Pixverse V6, Kling 3.0 Omni, VACE-14B, SkyReels-V3-14B, Phantom-14B, VINO-13B, Refalign-14B, Kaleido-14B, HunyuanCustom-13B, Bernini-14B, MV-S2V-14B, MAGREF-14B, and LTX-2-14B. Closed-source models were queried through official interfaces, open-source models used officially released weights with recommended settings, and all runs used 8 NVIDIA B200 GPUs.

Why This Matters

Impact on research. The paper argues that existing datasets and benchmarks (WebVid-10M, Panda-70M, InternVid, Cine250K, MuSS, OpenS2V-5M, ConsIDVid, and evaluation suites like VBench, VBench++, EvalCrafter, ChronoMagic-Bench, FETV, A2 Bench, OpenS2V-Eval, VACE-Bench, MultiShotMaster, MSAVBench) do not cover advertisement-specific requirements. AdSpark supplies both the training data and the measurement framework, which lets the field compare progress on product preservation, selling-point realization, and commercial effectiveness for the first time in a unified way.

Real-world applications:

• Automated production of e-commerce product advertisement videos from a single product image plus a written brief, replacing manual design and filming.
• Ad creative testing at scale, using AdSpark-Bench style metrics to rank many candidate videos before spending on media.
• Multi-shot product storytelling for short-form video platforms, where narrative pacing and shot transitions matter.
• Automated localized ad generation, since narration scripts and audio alignment are first-class annotations and metrics.

Industry relevance. The data originates from a major e-commerce platform with proper authorization, and the paper describes traditional ad production as labor-intensive and costly. A dataset of this size with structured commercial annotations, plus a benchmark that judges commercial persuasiveness, targets exactly the workflow that e-commerce and marketing teams pay for today. The authors state they will release the dataset upon acceptance.

Future Directions

• Improving product identity preservation under motion. The authors identify motion blur, shape distortion, and appearance drift as outstanding causes of degraded subject consistency, key-region similarity, and text fidelity.
• Better multi-shot narrative control in open models. Weak cross-shot product consistency and temporal coherence in open-source systems point to gaps the benchmark now makes measurable.
• Understanding data composition and scale. The paper runs controlled ablations on real-world versus synthetic data mix and training scale (Tab. 6), but the paper leaves the implications for scaling further open.
• Extending evaluation. The benchmark covers six dimensions; whether more dimensions are needed, and how reliable GPT-5.5-based judges are across them, remain open questions — the provided text notes metric validation exists in the supplementary material but does not report those results.

Target Audience

Researchers and engineers working on video generation, reference-conditioned and identity-preserving synthesis, and multimodal evaluation; applied scientists at e-commerce and marketing technology companies building automated ad creative pipelines; and benchmark designers who need a template for going beyond generic video quality metrics toward domain-specific, task-grounded evaluation. Readers should already be comfortable with diffusion video models and standard generation metrics.

Authors’ abstract

Product-centric advertisement video generation aims to create promotional videos that preserve fine-grained product identity while presenting selling points through coherent multi-shot narratives. However, this emerging task remains underexplored due to the lack of large-scale advertisement-specific datasets and comprehensive evaluation frameworks. To address this gap, we introduce \textbf{AdSpark}, a large-scale dataset and benchmark for product-centric advertisement video generation, based on data from a major e-commerce platform. \textit{AdSpark-300K} contains approximately 300K reference image--prompt--video triplets, comprising a real-world subset and a synthetic subset. Each sample provides structured advertisement annotations, including product identity annotations, selling-point descriptions, creative plans, and aligned audio scripts, enabling models to learn product preservation and advertisement-oriented visual storytelling. We further propose \textit{AdSpark-Bench}, a diagnostic benchmark that evaluates generated advertisements across six dimensions, including visual quality, product fidelity, instruction adherence, temporal coherence, audio alignment, and advertisement effectiveness. Based on AdSpark-Bench, we evaluate representative models, revealing key challenges in product preservation, multi-shot storytelling, and selling-point visualization. Experiments with AdSpark-300K-finetuned models further validate the effectiveness of our dataset. AdSpark provides a unified dataset and benchmark for future research, and we will release the dataset upon acceptance.