PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zhifei Yang, Zhao Jiang, Keyang Lu, Honghe Zhu, Zheng Zhang, Jingjing Lv, Changping Peng, Ching Law, Zhen Xiao

arXiv:2610.10047v1cs.CV

TL;DR

Product-centric advertisement generation must preserve fine-grained product identity while presenting selling points through coherent multi-shot narratives, yet suitable datasets and evaluation frameworks are lacking. AdSpark introduces a 300K-sample annotated dataset and a diagnostic benchmark; evaluations reveal model challenges and show improved advertisement-oriented generation after fine-tuning.

  • Problem

    Product-centric advertisement generation requires fine-grained product preservation and coherent selling-point visualization, but large-scale advertisement-specific datasets and benchmarks are lacking.

    Source: §1 · p. 1

  • Method

    AdSpark pairs 300K real-world and synthetic reference image–prompt–video samples with structured advertisement annotations and a benchmark for six dimensions of video and advertising capability.

    Source: §1 · pp. 2–3, §6 · p. 9

  • Results

    Evaluations reveal that existing models struggle with product-specific generation, while fine-tuning on AdSpark-300K improves advertisement capabilities.

    Source: §1 · p. 3, §6 · p. 9

  • Takeaways & Limitations

    AdSpark provides a unified dataset and benchmark for research on controllable, high-quality product-centric advertisement video generation.

    Source: §6 · p. 9

    The dataset currently draws on e-commerce scenarios and prioritizes products using click-through rate, leaving broader platforms, markets, styles, and long-tail products for future extension.

    Source: §D.1 · p. 35

Abstract

from arXiv · show

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.

1 INTRODUCTION

Product-centric advertisement video generation requires fine-grained product preservation and coherent selling-point storytelling, but existing datasets and benchmarks are not tailored to these needs. AdSpark pairs an annotated 300K-sample dataset with a diagnostic benchmark; evaluations reveal current model limitations and improvements after fine-tuning.

1 INTRODUCTION

  • Existing models struggle to preserve fine-grained product identity and visualize selling points through coherent multi-shot narratives, while existing datasets and benchmarks lack advertisement-specific coverage.General-purpose benchmarks emphasize visual quality, semantic alignment, and temporal coherence, but overlook product preservation, selling-point visualization, and commercial effectiveness.Source: §1 · pp. 1–2
  • AdSpark-300K contains 300K reference image–prompt–video triplets with structured annotations for product identity, selling points, creative plans, and aligned audio scripts.The dataset comprises 100K real-world and 200K synthetic videos, spanning 40 product categories and over 3,000 sub-categories.Source: §1 · pp. 2–3
  • AdSpark-Bench evaluates general video quality alongside advertisement-specific capabilities, including product fidelity, selling-point realization, and advertisement effectiveness.Its diagnostic scope also covers instruction adherence, temporal coherence, and audio alignment.Source: §1 · p. 3, Table 2 · p. 3
  • Evaluations reveal limitations in existing PC-AVG models, while fine-tuning a representative video generation model on AdSpark-300K substantially improves advertisement-oriented generation.The paper presents these evaluations as evidence of the dataset's effectiveness.Source: §1 · p. 3

2 RELATED WORK

Prior video datasets and benchmarks largely target general-purpose generation rather than product-centric advertising. AdSpark-300K and AdSpark-Bench address that mismatch with product-focused annotations and advertisement-specific evaluation.

2 RELATED WORK

  • Existing video datasets support generic text-to-video, multi-shot cinematic, or reference-based subject generation, but generally lack fine-grained product identity and commercial-generation annotations.AdSpark-300K instead covers product identities, selling points, creative plans, and audio scripts for product-centric advertisement generation.Source: §2 · p. 3
  • AdSpark-300K construction combines real-world advertisement curation with synthetic advertisement construction in tailored pipelines.The figure overview describes these as complementary branches with comprehensive, unified annotation.Source: Figure 2 · p. 4
  • Existing benchmarks assess general video quality, semantic alignment, temporal coherence, subject consistency, or cross-shot consistency, but overlook advertisement effectiveness and selling-point realization.AdSpark-Bench provides a dedicated evaluation framework for PC-AVG.Source: §2 · pp. 3–4

3 ADSPARK-300K

AdSpark-300K combines real-world and synthetic advertisement data through tailored curation and construction pipelines that produce a shared annotated format. Its statistics show diverse narration and duration patterns, with a predominantly multi-shot structure.

3 ADSPARK-300K

  • AdSpark-300K combines complementary real-world and synthetic subsets through multi-stage filtering, annotation, and quality control, using a unified reference-image, video, and annotation format.The paired assets and associated product metadata come from an e-commerce platform with proper authorization.Source: §3 · p. 4

3.1 REAL-WORLD ADVERTISEMENT CURATION

  • Real-world curation segments and tracks products, filters frames and clips for visibility and identity consistency, annotates shots and advertisement content, then manually verifies samples.Filtering uses geometric and temporal checks followed by VLM review; annotations include product identity, selling points, creative plans, and shot-aligned audio transcripts.Source: §3.1 · pp. 4–5

3.2 SYNTHETIC ADVERTISEMENT CONSTRUCTION

  • The synthetic subset filters product SKUs, generates and reviews structured advertisement plans, synthesizes videos with multiple models, and removes outputs that fail human quality inspection.Plans translate selling points into visualizable actions and specify shot-level timing, camera motion, and shot type; accepted samples share the unified annotation format.Source: §3.2 · p. 5

3.3 DATA ANALYSIS

  • Over 85% of samples contain multiple shots, primarily 2-shot (47.3%) and 3-shot (23.6%) compositions; the dataset also spans diverse narration paces and shows high visual quality by MUSIQ.Figure 3 summarizes duration, audio density, shot-count, and normalized MUSIQ score distributions.Source: §3.3 · pp. 5–6, Figure 3 · p. 5

4 ADSPARK-BENCH

AdSpark-Bench is a held-out, diagnostic benchmark for product-centric advertisement video generation, with six evaluation dimensions and shot-aware assessment. Its test cases cover diverse product categories and include multi-shot advertisements for evaluating cross-shot behavior.

4 ADSPARK-BENCH

  • AdSpark-Bench evaluates generated advertisements across visual quality, product fidelity, instruction adherence, temporal coherence, audio alignment, and advertisement effectiveness.Its hierarchical framework is tailored to product-centric advertisement video generation.Source: §4 · p. 6

4.1 BENCHMARK STATISTICS

  • The 220 SKU-disjoint test cases span all 40 product categories and 220 sub-categories, including 175 multi-shot cases.The set comprises 45 one-shot, 131 two-shot, and 44 three-shot cases; all are manually reviewed.Source: §4.1 · p. 6

4.2 HIERARCHICAL EVALUATION METRICS

  • Visual quality uses Aesthetic Score and MUSIQ, while product fidelity separately measures full-product appearance, identity-defining regions, product text, and consistency across shots.Foreground regions are extracted at different granularities to reduce sensitivity to background variation.Source: §4.2 · pp. 6–7
  • Instruction adherence assesses planned shot structure, shot-level execution, and scene, style, and selling-point realization, alongside global video–text alignment.Shot execution uses sampled frames and optical-flow motion cues; selling-point assessment uses denser shot-aware sampling.Source: §4.2 · p. 7
  • Temporal coherence covers within-shot motion, transition naturalness, and stability around shot boundaries, while audio alignment checks background sound and narration against the plan.Advertisement effectiveness assesses attractiveness, creative quality, and narrative coherence from a potential-customer perspective.Source: §4.2 · p. 7
  • Table 3 compares proprietary, open-source, and AdSpark-finetuned models; unavailable or invalid audio outputs are excluded from audio-alignment evaluation.Bold and underlined entries mark the best and second-best results, respectively.Source: Table 3 · p. 8

5 EXPERIMENTS

Experiments compare closed- and open-source video generation models on AdSpark-Bench and fine-tune LTX-2 on AdSpark-300K. Existing models show advertisement-specific weaknesses, while the fine-tuned model improves several cross-shot and narrative measures over its base model.

5.1 EXPERIMENTAL SETTINGS

  • The evaluation covers closed-source and open-source models on AdSpark-Bench; LTX-2 is additionally fine-tuned on AdSpark-300K with rank-64 LoRA as LTX-AdSpark.Closed-source models use official interfaces, while open-source models use released weights and recommended inference settings.Source: §5.1 · p. 8

5.2 EVALUATION RESULTS

  • 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 over LTX-2.The authors report these gains as evidence of AdSpark-300K's utility for product-centric advertisement video generation.Source: §5.2 · p. 8
  • Closed-source models generally outperform open-source models, but product fidelity remains challenging during complex motion, with blur, shape distortion, and appearance drift.Open-source models show more pronounced limitations in instruction adherence, advertisement effectiveness, multi-shot generation, and temporal coherence.Source: §5.2 · p. 8
  • Qualitative comparisons show color shifts, product deformation, and instruction deviations; LTX-AdSpark appears more product-consistent and coherent across shots than LTX-2.All methods are conditioned on the corresponding reference image and advertisement prompt.Source: §5.2 · p. 9, Figure 4 · p. 9

6 CONCLUSION

The paper concludes that AdSpark combines a large-scale annotated dataset with a six-dimension benchmark for product-centric advertisement video generation. Experiments find challenges in existing models and improvements from AdSpark-300K fine-tuning.

6 CONCLUSION

  • AdSpark-300K provides 300K reference image–prompt–video triplets with structured advertisement annotations, while AdSpark-Bench evaluates six dimensions.The benchmark dimensions are visual quality, product fidelity, instruction adherence, temporal coherence, audio alignment, and advertisement effectiveness.Source: §6 · p. 9
  • Experiments show that existing models struggle with product-specific generation, while fine-tuning on AdSpark-300K improves advertisement capabilities.Source: §6 · p. 9
  • The authors hope AdSpark will facilitate future research on controllable and high-quality product-centric advertisement video generation.Source: §6 · p. 9

Supplementary Material Overview. The supplementary material is organized as follows:

The supplementary material details AdSpark’s data sources and curation, synthetic-ad construction, and unified annotations, alongside benchmark, experimental, and broader-impact material. The described pipelines turn e-commerce product assets into quality-controlled real-world and synthetic advertisements with structured supervision.

A.1 DATA SOURCES AND COLLECTION PROTOCOL

  • AdSpark-300K draws on authorized internal e-commerce data; real-world samples require linked video, reference image, and metadata, while synthetic samples need an image and metadata.The benchmark is SKU-disjoint from AdSpark-300K, although the real-world and synthetic subsets may share SKUs.Source: §A.1 · p. 14

A.2 REAL-WORLD ADVERTISEMENT CURATION

  • 192,077 SKUs yielded 218,837 eligible videos; tracking and frame filters produce 5–20-second clips that are quality-screened, shot-segmented, annotated, and manually verified.Tracking uses the reference-image mask; clips are retained only after product consistency, visibility, and annotation checks.Source: §A.2 · pp. 14–15

A.3 SYNTHETIC ADVERTISEMENT CONSTRUCTION

  • Synthetic candidates pass asset-quality, generation-suitability, and category filters; 348,526 remaining SKUs are click-ranked until 200K quality-controlled advertisements are collected.Asset checks include valid metadata and transparent-background images; suitability and category filters exclude unsuitable subjects, content, and products.Source: §A.3 · pp. 15–16
  • GPT-5.5 plans one to five ordered shots from product metadata and selling points, while Gemini reviews annotations for accuracy, coherence, plausibility, and visual–audio alignment.Plans specify shot intervals, types, camera motion, visual instructions, and aligned audio instructions; annotations are revised for up to five review rounds.Source: §A.3 · pp. 16–17
  • Synthetic videos use Seedance 2.0, HappyHorse, or Kling to generate 5–15-second vertical 9:16 videos with audio, without promotional overlays or other non-product text.Seedance 2.0 generates most of the synthetic subset; outputs are 720p and include narration, background sound, and action-related effects.Source: §A.3 · p. 17
  • Each synthetic ad is manually checked for product fidelity, plausibility, temporal quality, visible selling-point realization, and audio quality and alignment.Examples of rejection criteria include identity drift, implausible interactions, incoherent transitions, omitted or contradicted selling points, and audio–visual mismatch.Source: §A.3 · p. 17

A.4 UNIFIED ANNOTATION

  • Both subsets use a shared representation covering product identity, selling points, creative plans, aligned audio scripts, and generation prompts.Identity fields capture appearance and key regions; selling-point annotations specify visual realization and observable success criteria, while plans connect ordered shots to selling points.Source: §A.4 · pp. 17–18
  • Real-world annotations describe source-ad content, whereas synthetic annotations are planned before generation; fields are normalized and checked for completeness, timing, references, and consistency.The unified representation supports reference-conditioned generation, identity preservation, selling-point visualization, multi-shot planning, and audio–visual generation.Source: §A.4 · pp. 18–19

A.5 DATASET STATISTICS AND DISTRIBUTIONS

AdSpark-300K’s content distributions show a preference for smooth product-focused camera motion, demonstration-oriented shots, and advertisements centered on a compact set of selling points.

A.5 DATASET STATISTICS AND DISTRIBUTIONS

  • Tracking shots account for 38.1% of camera motions, dolly-ins 35.2%, and rack focus 13.6%.The distribution favors smooth subject following and gradual visual emphasis.Source: §A.5 · p. 19
  • Demonstration shots are most frequent, followed by hero, macrodetail, and full-product shots.The distribution reflects the importance of functional visualization and fine-grained product presentation.Source: §A.5 · p. 19
  • Most products are associated with two or three selling points; advertisements with more than three are relatively rare.The dataset generally focuses on a compact set of core commercial messages.Source: §A.5 · p. 19
  • Prompt lengths peak around 500 characters while spanning a wide range.The range reflects variation in scene detail, shot complexity, and selling-point presentation.Source: §A.5 · p. 19

A.6 REPRESENTATIVE SAMPLES

This section presents qualitative AdSpark-300K examples and describes AdSpark-Bench’s construction, coverage, and evaluation protocols across visual quality, product fidelity, instruction adherence, temporal coherence, audio alignment, and advertisement effectiveness. It also specifies multimodal judging and score normalization, defining metric scope and conditions rather than reporting model results.

A.6 REPRESENTATIVE SAMPLES

  • Qualitative samples show product identity and selling points preserved across diverse categories, scenes, camera motions, and interactions, alongside structured multi-shot prompts and synchronized audio cues.The examples include scene and style descriptions, shot-level composition and camera motion, temporal segmentation, and audio scripts.Source: §A.6 · pp. 19–20, Figure 7 · p. 20, Figure 8 · p. 21

B.1 BENCHMARK CONSTRUCTION

  • The benchmark’s final suite contains 220 product-conditioned cases spanning 40 categories and 220 distinct sub-categories, selected from 350 SKU-disjoint candidates.Filtering requires clear product identity, high-quality images, complete annotations, and one sample per fine-grained category; 175 cases are multi-shot.Source: §B.1 · pp. 21–22

B.2 CATEGORY AND DIFFICULTY DISTRIBUTION

  • The benchmark covers all 40 dataset categories and balances shot complexity, combining product-detail, packaging-text, logo, storytelling, and product-motion challenges.One-shot cases assess basic preservation; two- and three-shot cases add transition, temporal-consistency, and narrative-progression challenges.Source: §B.2 · p. 22

B.3 CONDITIONAL EVALUATION SUBSETS

  • Conditional evaluation uses subsets of 72 text-fidelity, 172 logo/key-region, and 20 product-demonstration samples, plus 175 multi-shot cases for multi-shot metrics.These subsets supply the visual prerequisites needed for reliable metric measurements.Source: §B.3 · p. 22

B.4 SHOT-AWARE EVALUATION PROTOCOL

  • A shot-aware protocol matches generated shots to planned shots using temporal IoU, then assesses shot-structure alignment, shot execution, within-shot motion, and transitions without ground-truth boundaries.Transition metrics are skipped for single-shot advertisements; other applicable metrics remain unchanged.Source: §B.4 · pp. 22–23

B.5.1 VISUAL QUALITY

  • Visual quality is evaluated without considering product identity or instruction compliance using Aesthetic Score and MUSIQ; for N = 16 sampled frames, frame-level scores are independently computed and averaged.Aesthetic Score is normalized to [0, 1], while MUSIQ directly produces a normalized image quality score.Source: §B.5.1 · p. 23

B.5.2 PRODUCT FIDELITY

  • Product fidelity measures overall appearance, identity-defining regions, packaging text, and cross-shot stability using product-region features, annotated key regions, OCR, and shot transitions.The protocol uses GroundingDINO and SAM2 to extract product regions and DINOv3 for feature comparisons.Source: §B.5.2 · pp. 23–24

B.5.3 INSTRUCTION ADHERENCE

  • Instruction adherence evaluates shot structure, shot-level execution, scene and style requirements, and selling-point realization through shot-aware metrics reported independently.GPT-5.5 evaluates matched-shot execution and content alignment; GmeScore supplies complementary global prompt-video alignment.Source: §B.5.3 · pp. 25–26

B.5.4 TEMPORAL COHERENCE

  • Temporal coherence combines product-motion stability, transition naturalness, and boundary-level temporal consistency, with transition metrics limited to videos containing detected shot boundaries.Product stability tracks centroid, area, and mask-IoU stability; transition naturalness checks corrupted frames and judged visual continuity, while temporal consistency detects flicker and structural artifacts.Source: §B.5.4 · pp. 27–28

B.5.5 AUDIO ALIGNMENT

  • Audio alignment is evaluated only for valid audio outputs through independent background-sound similarity and narration-script consistency metrics.LAION-CLAP compares generated audio with the planned sound description, while Whisper transcription is compared with the reference script using normalized edit distance; samples without narration are not applicable.Source: §B.5.5 · p. 28

B.5.6 ADVERTISEMENT EFFECTIVENESS

  • Advertisement effectiveness uses a potential-customer perspective to score attractiveness, creative quality, and narrative coherence independently, with coherence criteria adapted to single- or multi-shot ads.The GPT-5.5 evaluator considers sampled video frames alongside the reference product, creative plan, and selling points; multi-shot coherence assesses logical connections and selling-point progression.Source: §B.5.6 · pp. 28–29

B.6 MULTIMODAL JUDGE PROMPTS

  • GPT-5.5 judges selected advertisement-specific metrics using ordered frames or contact sheets with structured annotations, and mapped Likert scores or direct scalar scores.The protocol requires strict JSON or numeric outputs; evaluation uses independently generated visual evidence under identical reference-image and prompt conditioning, with metrics validated against human judgments.Source: §B.6 · pp. 29–31

B.7 SCORE NORMALIZATION AND AGGREGATION

  • Scores are normalized to [0, 1] before percentage reporting, and aggregated only over valid units and applicable benchmark samples.Conditional metrics exclude samples lacking evaluation targets rather than assigning zero, except planned multi-shot cases that fail to produce multiple detected shots receive zero for the corresponding cross-shot metric.Source: §B.7 · pp. 31–32

B.8 METRIC VALIDATION

The authors validate six GPT-assisted metrics against human rankings, finding positive agreement across dimensions that is stronger for objective criteria than subjective ones. Benchmark examples show how structured cases support evaluation across 40 product categories.

B.8 METRIC VALIDATION

  • The validation compares GPT-5.5 scores with blinded human ratings on videos from 19 models generated for 100 held-out products.Source: §B.8 · p. 32
  • GPT-assisted metrics correlate positively with human rankings across all six dimensions, with stronger agreement on objective criteria than on creative quality and advertisement attractiveness.Source: §B.8 · p. 33, Table 5 · p. 33

B.9 BENCHMARK EXAMPLES

  • Benchmark examples cover 40 product categories and pair product identity and selling points with shot plans, audio scripts, and generation prompts.Source: §B.9 · p. 33

C.1 DETAILS OF EVALUATION MODELS

The evaluation covers proprietary and open-source reference-to-video models under shared product references and prompts where supported. Qualitative comparisons show that LTX-AdSpark produces more product-consistent videos with smoother shot progression and clearer selling-point presentation.

C.1 DETAILS OF EVALUATION MODELS

  • The evaluation covers proprietary and open-source models, using the same product references and prompts when supported; audio metrics are unavailable for models without audio generation.Source: §C.1 · pp. 33–34

C.2 IMPLEMENTATION DETAILS

  • LTX-AdSpark adapts LTX-2 using LoRA on attention projections, with product images as visual conditioning and structured advertisement prompts as textual guidance.Source: §C.2 · p. 34

C.3 QUALITATIVE COMPARISONS

  • LTX-AdSpark produces more product-consistent videos, smoother shot progression, and clearer selling-point presentation, while other models show product distortions and weak shot-level prompt following.Source: §C.3 · p. 34

C.4 ABLATION RESULTS

The ablations find that real-world and synthetic data contribute complementary strengths, and the full AdSpark-300K mix performs best overall. Increasing synthetic data scale also improves all reported metrics.

C.4 ABLATION RESULTS

  • 65.01 vs. 61.59 Subject Consistency follows adding 100K real-world samples to Synthetic-200K, and the full AdSpark-300K performs best overall.The compared variants share the LTX-2 initialization and training settings unless otherwise specified.Source: §C.4 · pp. 34–35, Table 6 · p. 35
  • 64.73 vs. 58.17 Subject Consistency favors Real-100K over Synthetic-100K, while Synthetic-100K scores higher on Shot Structure Alignment, Transition Naturalness, and Narration Script Consistency.Synthetic-100K scores 71.79, 80.09, and 81.38 on those three metrics, versus 70.68, 72.76, and 78.26 for Real-100K.Source: §C.4 · p. 35
  • Scaling the synthetic subset from 100K to 200K brings further gains across all reported metrics, and both real-world and synthetic subsets improve over the base model.Source: §C.4 · p. 35

D.1 LIMITATIONS AND FUTURE WORK

AdSpark is currently constructed from e-commerce scenarios, and product selection is mainly guided by click-through rate. The authors identify broader sources and long-tail product coverage as future extensions.

D.1 LIMITATIONS AND FUTURE WORK

  • The dataset currently draws from e-commerce scenarios; extending to other platforms, markets, and advertising styles is identified as future work.Source: §D.1 · p. 35
  • Click-through-rate-guided selection prioritizes representative, commercially relevant products, while broader long-tail coverage remains a future extension.Source: §D.1 · p. 35
  • The proposed broader coverage would include niche categories and less frequently promoted products to support more comprehensive evaluation and generation across product types.Source: §D.1 · p. 35
Loading 2610.10047v1…