PaperlayerView exact revision on arXiv

Source-linked AI summary

Co-Director: Agentic Generative Video Storytelling

Yale Song, Yiwen Song, Nick Losier, Nathan Hodson, Ye Jin, Rhyard Zhu, Yan Xu, Daniel Vlasic, Carina Claassen, Jasmine Leon, Khanh G. LeViet, Zack Chomyn, Joe Timmons, Brett Slatkin, Scott Penberthy, Tomas Pfister

arXiv:2604.24842v1cs.AIcs.MAcs.MM

TL;DR

Co-Director addresses the challenge of turning generated video clips into coherent end-to-end stories without semantic drift. Its hierarchical multi-agent framework combines global creative-direction optimization with local multimodal refinement, outperforming monolithic models and existing agentic pipelines while generalizing beyond advertising to cinematic narratives.

  • Problem

    Existing video-storytelling benchmarks evaluate isolated subtasks, leaving end-to-end transformation of abstract ideas into cohesive narratives insufficiently assessed.

  • Method

    Co-Director combines multi-armed-bandit steering of global creative intent with a local multimodal feedback loop for coherent multi-agent video storytelling.

  • Results

    Co-Director outperforms monolithic models and existing agentic pipelines, with qualitative evidence indicating consistency and creative steering generalize beyond advertising to cinematic narratives.

  • Takeaways & Limitations

    The framework offers a principled approach to autonomous visual storytelling under strict targeting and visual constraints.

  • Takeaways & Limitations

    The MLLM-based evaluation metrics may introduce bias concerns.

Abstract

from arXiv · show

While diffusion models generate high-fidelity video clips, transforming them into coherent storytelling engines remains challenging. Current agentic pipelines automate this via chained modules but suffer from semantic drift and cascading failures due to independent, handcrafted prompting. We present Co-Director, a hierarchical multi-agent framework formalizing video storytelling as a global optimization problem. To ensure semantic coherence, we introduce hierarchical parameterization: a multi-armed bandit globally identifies promising creative directions, while a local multimodal self-refinement loop mitigates identity drift and ensures sequence-level consistency. This balances the exploration of novel narrative strategies with the exploitation of effective creative configurations. For evaluation, we introduce GenAD-Bench, a 400-scenario dataset of fictional products for personalized advertising. Experiments demonstrate that Co-Director significantly outperforms state-of-the-art baselines, offering a principled approach that seamlessly generalizes to broader cinematic narratives. Project Page: https://co-director-agent.github.io/

1. Introduction

Co-Director frames generative video storytelling as a global optimization problem and addresses semantic drift, cascading failures, and limited creative exploration in existing agentic pipelines. It combines hierarchical creative control with bandit optimization and local multimodal refinement, evaluated on the 400-scenario GenAd-Bench advertising dataset.

  • Motivation and Challenges: Existing agentic pipelines remain vulnerable to cascading failures, semantic drift, and fixed handcrafted prompts that hinder long-horizon consistency and creative exploration.Early errors can propagate through chained generation, while static templates cannot adapt to non-linear creative processes.
  • Co-Director: Co-Director formalizes generative storytelling as a global optimization problem using hierarchical parameterization steered by a Multi-Armed Bandit.The framework replaces linear waterfall chains and fixed prompt templates with a hierarchical strategy for creative direction.
  • Co-Director: Co-Director integrates multi-armed bandit optimization with a local MLLM feedback-driven refinement loop to achieve global coherence.The architecture is designed to coordinate visual and narrative consistency across the storytelling process.
  • Evaluation and Results: GenAd-Bench evaluates generative video storytelling across 400 scenarios using fictional brand assets, while Co-Director outperforms foundation models and agentic baselines in asset preservation, contextual alignment, and visual quality.Digital advertising provides a constraint-driven setting requiring asset preservation, audience alignment, and value proposition communication.

2. Related Work

Related work shows that video synthesis models can achieve cinematic realism and improve character consistency, yet remain limited as standalone tools for long-range narrative coherence and high-level planning. Agentic storytelling frameworks address clip-to-narrative decomposition but face local consistency and global credit-assignment challenges, while existing benchmarks do not fully capture professional visual and narrative requirements.

  • Foundation Models for Video Synthesis: Large-scale diffusion models achieve unprecedented cinematic realism, but maintaining long-range consistency remains challenging.
  • Foundation Models for Video Synthesis: Phantom and MoC introduce advanced character-consistency techniques, but remain standalone generators lacking high-level planning and reasoning for complex narratives.
  • Agentic Frameworks for Video Storytelling: Recent multi-agent systems decompose scripts into shot lists and prompts for personalized vlogs, music videos, and short films.
  • Agentic Frameworks for Video Storytelling: AniMaker and MAViS incorporate local consistency checks, but struggle with global credit assignment across the agentic stack.
  • Agentic Frameworks for Video Storytelling: Co-Director introduces a Multi-Armed Bandit formulation for principled exploration-exploitation of the creative latent space, addressing semantic collisions between sub-agents.
  • Benchmarks for Video Storytelling: Generic video benchmarks focus on prompt alignment and broad aesthetics, lacking the visual and narrative qualities demanded by professional workflows.

3. Approach

Co-Director orchestrates video generation as a hierarchical multi-agent pipeline, using global creative-direction optimization and local artifact refinement to improve narrative and visual consistency. The pipeline plans storylines and storyboards, generates keyframes, video, and audio, then assembles the final video.

  • Global orchestration: The Orchestrator Agent parses prompts and reference visuals into structured constraints, samples creative configurations with a multi-armed bandit, and propagates global conditioning across the pipeline.At iteration t, it forms Θ(t) from structured user constraints, annotated references, and the sampled configuration θ(t).
  • Pre-production: The Pre-Production Agent converts constraints into a creative brief, scene-by-scene storyline, missing visual assets, and a storyboard specifying semantics, camera motion, timing, entities, and audio directives.Visual assets are generated only when absent from reference visuals, while the storyboard expands the storyline into scenes s_i.
  • Production and assembly: The Production Agent generates keyframes as visual priors, conditions video diffusion on each keyframe, and synthesizes global voiceover and music before post-production combines clips and audio.Keyframes mitigate identity drift, while audio properties such as tempo, genre, and mood are conditioned on θ.
  • Local refinement: Feedback descent locally refines storylines and keyframe sequences through iterative LLM or MLLM evaluation, textual feedback, and regeneration when scores fall below thresholds.Storylines are assessed for narrative and advertising qualities, while keyframes are jointly evaluated for identity and environmental continuity.
  • Hierarchical parameterization: Hierarchical parameterization treats the non-differentiable pipeline as a black box and tunes global creative directions θ = (θ_cs, θ_nm, θ_aa) through a multi-armed bandit.The three axes are creative strategy, narrative mode, and aesthetic archetype, balancing exploration of untried directions with exploitation of high-reward options.

4. GenAd-Bench: Evaluating Generative Video Advertising

GenAd-Bench is a 400-scenario benchmark of fictional products that tests targeted, multimodally conditioned video advertising across paired demographics and geographic settings. Generated videos are scored by an MLLM on four 0–100 dimensions, with human-study alignment supporting its reliability as an evaluation proxy.

  • Benchmark Construction: GenAd-Bench contains 400 unique scenarios built from 50 fictional brands with four products each and paired stereotypical and unconventional demographics.Each scenario is defined by brand, product, gender, age, location, and interest.
  • Benchmark Design: The benchmark requires models to combine textual demographic intent with specific reference imagery under strict targeting constraints.Paired scenarios across global geographic settings test conceptual flexibility and exposure to ingrained biases.
  • Benchmark Design: Using fictional entities prevents reliance on memorized training priors, making benchmark performance measure reasoning rather than retrieval.The paired-scenario design also evaluates flexibility across demographic and geographic variations.
  • Evaluation Protocol: Gemini 3 Pro evaluates generated videos by jointly processing visual frames and audio across four dimensions scored from 0–100.The dimensions are Visual Asset Fidelity, Demographic Alignment, Marketing Appeal, and Visual Quality.
  • Evaluation Protocol: The metrics assess identity preservation, demographic targeting, advertising persuasiveness, and broadcast-oriented visual quality.Visual Quality penalizes diffusion artifacts, object morphing, temporal flickering, unnatural motion, and violations of basic physical laws.
  • Evaluator Validation: Human-study comparison shows that the MLLM evaluator aligns closely with human judgments and approaches the human-human agreement ceiling.This comparison addresses potential bias concerns associated with MLLM-based evaluation.

5. Experiments

Co-Director is evaluated on GenAd-Bench against external and agentic baselines, with ablations showing that global MAB optimization and local self-refinement are necessary for performance. Human-study and qualitative analyses further assess metric alignment and generalization to cinematic narratives.

  • Overall evaluation: Co-Director achieves the highest average performance on GenAd-Bench, scoring 81.4 across scaled visual, demographic, marketing, and quality metrics.Table 1 evaluates Visual Asset Fidelity, Demographic Alignment, Marketing Appeal, and Visual Quality on a [0, 100] scale.
  • Baseline comparisons: Using identical foundation models, the Base Agentic Pipeline averages 68.5, while AniMaker and MovieAgent score 62.2 and 65.3, respectively.The comparison isolates Co-Director’s orchestration from differences in underlying models.
  • Optimization efficiency: Warm-Up + MAB reaches high-quality results in the first iteration, whereas Random Search improves more steadily and eventually discovers strong configurations through undirected sampling.The cumulative-best comparison is reported over 10 iterations.
  • Metric validation: On a 50-scenario human study with 5 raters per video, MLLM judgments approach human agreement on complex dimensions such as DA and MA, while VQ correlates less strongly.The study evaluates VAF, DA, MA, and VQ using a 5-point Likert scale and reports Krippendorff’s Alpha, Cohen’s Kappa, Pearson, Spearman, and MAE.
  • Ablation studies: Collapsing arm-specific rewards into one scalar reduces VAF by 5.0 points, and removing all refinement and optimization produces a 67.2 Avg. barebones baseline.The complete architecture outperforms this linear performance floor by a relative 17.6%.
  • Ablation studies: Removing Keyframe Refinement causes a 9.8-point VAF drop, while removing Storyline Refinement primarily harms MA and DA.The ablation identifies keyframe refinement as vital for product identity and storyline refinement as important for narrative coherence.

6. Conclusion · A. Appendix Overview

Co-Director formulates video storytelling as a global optimization problem and uses MAB-driven steering to achieve high-fidelity narrative consistency. The appendix organizes qualitative results, benchmark evaluations, dataset details, and baseline implementation information.

  • 6. Conclusion: Co-Director is presented as a hierarchical multi-agent framework for video storytelling.The framework formulates video storytelling as a global optimization problem.
  • 6. Conclusion: MAB-driven steering navigates a creative latent space to achieve high-fidelity narrative consistency.The conclusion states that this consistency traditionally requires large-scale production teams.
  • 6. Conclusion: GenAd-Bench evaluates end-to-end generative workflows under strict targeting and visual constraints.The conclusion describes GenAd-Bench as a rigorous dataset designed for this evaluation.
  • A. Appendix Overview: The appendix overview serves as a directory for supplemental materials covering results, dataset, and methodology components.Each appendix section covers a distinct component of the paper’s results, dataset, or methodology.
  • A. Appendix Overview: Section B presents qualitative GenAd-Bench results through visual galleries of brand-identity preservation across diverse narrative arcs.These galleries demonstrate Co-Director’s ability to preserve brand identity.
  • A. Appendix Overview: Section C evaluates general video storytelling using performance metrics and qualitative comparisons on ViStoryBench.The benchmark is attributed to Zhuang et al., 2025.
  • A. Appendix Overview: Section D documents GenAd-Bench dataset construction and descriptive statistics.This section provides details on how the dataset was constructed and characterized.
  • A. Appendix Overview: Section E describes baseline implementation details through standardized configurations and prompt settings.The supplied passage lists standardized configurations and prompt temperature information before truncation.

B. Qualitative Results on GenAd-Bench

Co-Director translates six-point product prompts and reference visuals into cohesive, high-fidelity visual narratives on GenAD-Bench. Qualitative examples show semantic grounding, identity preservation, and consistency across varied scenarios, settings, poses, and camera angles.

  • Input-to-narrative translation: Co-Director converts six-point product prompts, brand logos, and product references into cohesive, high-fidelity visual narratives.The qualitative evaluation focuses on the framework’s ability to translate these inputs into visual storytelling.
  • Storyline coherence: Figure 3 demonstrates precise scene instructions, rich emotional arcs, and robust character and product identity preservation across four scenarios.The examples ground abstract marketing concepts, including a functional bellows effect and a psychological shift toward safety, in synthesized storylines and keyframes.
  • Scenario versatility: Figure 4 shows consistent atmosphere, structure, and temporal continuity across 10 additional stereotypical and unconventional product scenarios.The examples include appropriate rustic or softly lit settings and correct object manipulation in unexpected spaces such as high-tech workspaces.

C. Evaluation on General Video Storytelling · D. GenAd-Bench Details

Co-Director is evaluated as a general video-storytelling framework beyond advertising, including an ablated ViStoryBench-Lite setting and qualitative analyses. GenAd-Bench is introduced to rigorously evaluate video storytelling through dataset construction, demographic and category statistics, data splits, and qualitative examples.

  • C. Evaluation on General Video Storytelling: Video advertising demonstrates Co-Director’s ability to generate cohesive narratives from a single high-level idea, unlike AniMaker and MovieAgent, which require scene-by-scene scripts.The paper selects advertising because it tests full narrative generation rather than execution of a detailed script.
  • C. Evaluation on General Video Storytelling: Co-Director’s generality is tested on ViStoryBench-Lite by bypassing script generation and video synthesis to match its storyboard-generation inputs and static-image outputs.The ablation addresses the benchmark’s architectural mismatch with Co-Director’s end-to-end pipeline.
  • C. Evaluation on General Video Storytelling: Co-Director records the highest Style consistency scores, with Cross: 0.499 and Self: 0.743, and the highest Character consistency scores, with Cross: 0.499 and Self: 0.593.These results indicate strong visual consistency relative to prior agentic baselines.
  • C. Evaluation on General Video Storytelling: Co-Director achieves the highest average Prompt Alignment score of 3.166, while trailing in Quality & Diversity, including an Inception Score of 10.900.The prompt-alignment result reflects control over camera dynamics and local context, whereas the reported quality-diversity example is lower.
  • C. Evaluation on General Video Storytelling: Qualitative results show cohesive narratives that preserve character and product identity across dynamic poses, varied angles, environmental transitions, and lighting changes.These observations are illustrated on GenAd-Bench examples spanning complex product prompts and dramatic visual changes.
  • C. Evaluation on General Video Storytelling: Across diverse narratives, Co-Director maintains temporal coherence, character identity, and background spatial structures, with strong prompt alignment and character preservation despite minor anatomical artifacts.The paper uses Figures 5 and 6 to corroborate the quantitative metrics with sequence-level observations.
  • D. GenAd-Bench Details: GenAd-Bench is introduced as a rigorous video-storytelling evaluation dataset covering construction, demographic and category statistics, data splits, and qualitative examples.The appendix presents these components to document the benchmark comprehensively.

D.1. Dataset Construction Pipeline · D.2. Taxonomy and Distribution Statistics

GenAd-Bench uses a human-in-the-loop pipeline to create fictitious branded products and paired advertising scenarios, with visual assets designed for brand accuracy. Its 400 scenarios span diverse categories and enforce balanced demographic and geographic distributions.

  • D.1. Dataset Construction Pipeline: GenAd-Bench synthesizes 200 fictitious products across 50 fictitious brands through a human-in-the-loop pipeline.Each brand contributes four products, creating an evaluation setting that reduces reliance on parametric memory.
  • D.1. Dataset Construction Pipeline: The construction pipeline uses Gemini 3 Pro for structural text metadata and Nano Banana Pro for visual reference assets within a hierarchical schema.These models support separate textual and visual stages of dataset construction.
  • D.1. Dataset Construction Pipeline: The pipeline assigns exactly four distinct products to each brand and spans consumer goods, industrial equipment, software, and subscription services.This broad form-factor coverage tests generation beyond a narrow class of physical objects.
  • D.1. Dataset Construction Pipeline: Each product receives contrasting Stereotypical and Unconventional scenarios to test conceptual flexibility across traditional and expectation-subverting target demographics.The paired design varies demographic framing while keeping the underlying product fixed.
  • D.1. Dataset Construction Pipeline: Distinct vector-style brand logos and product reference images are generated conditionally to integrate textual product descriptions with brand identity.The targeted refinement step ensures uniqueness and brand accuracy before assets propagate downstream.
  • D.2. Taxonomy and Distribution Statistics: 400 evaluation scenarios are distributed across diverse industries and global contexts.The scenarios derive from the 200 products and their paired demographic-targeting configurations.
  • D.2. Taxonomy and Distribution Statistics: The 200 products span 24 macro-categories, including high-frequency Home Goods (24), Food (20), Fashion (20), and Tech (20).Specialized domains include Automotive (8), Logistics (4), and Baby (4), probing niche context synthesis.
  • D.2. Taxonomy and Distribution Statistics: The dataset maintains exact gender parity: 50% (200) of target personas are Male and 50% (200) are Female, while paired scenarios neutralize product-level gender bias.Balanced demographic and geographic distributions are intended to prevent optimization from exploiting ingrained training biases.

D.3. Dataset Split

GenAd-Bench is split into equal Hillclimbing and Validation sets, with all reported metrics evaluated on Validation across in-domain and out-of-domain challenges spanning 45 brands.

  • Dataset Split: The 400 scenarios are partitioned into two equal 200-scenario splits: Hillclimbing for iterative optimization and Validation for evaluation.The Hillclimbing Set is excluded from the reported evaluations.
  • Dataset Split: Validation is stratified into two sub-challenges assessing generalization across 45 unique brands.Metrics are reported solely on the Validation Set.
  • In-Domain: In-Domain comprises 80% of Validation, or 160 scenarios, covering 80 products from 40 brands sharing macro-categories with Hillclimbing.It tests extrapolation to novel products and brands within familiar industries.
  • Out-of-Domain: Out-of-Domain comprises 20% of Validation, or 40 scenarios, covering 20 products from 5 brands in industries absent from Hillclimbing.It tests generalization to product categories without category-specific optimization.

E. Baseline Implementation Details … G.2. Visual Realization via Aesthetic Archetypes

The evaluation compares standardized zero-shot monolithic/commercial systems and structured-script agentic pipelines, while Co-Director uses sample-efficient MAB optimization and hierarchical creative direction. Across aesthetic archetypes, this steering changes visual style while preserving reference assets, logos, and structural features.

  • E. Baseline Implementation Details: Monolithic and commercial baselines are evaluated with a standardized zero-shot prompting strategy.Systems include LTX-2.3, Kling 3.0 Omni, Veo 3.1, Wan 2.6, Creatify, and HeyGen.
  • E. Baseline Implementation Details: AniMaker and MovieAgent are compared using a Gemini 3 Pro Script Generator that converts product briefs into structured multi-shot scripts.This isolates multi-agent architectural efficacy from differences in underlying generative backbones.
  • F. Extended Discussion on MAB Optimization Efficiency: 36 end-to-end video-generation pipelines would be required to exhaustively search the creative action space for each prompt.The brute-force approach guarantees the global optimum but is computationally and financially unscalable.
  • F. Extended Discussion on MAB Optimization Efficiency: Factored rewards update three creative axes simultaneously, reducing the search from 36 paths to 10 independent parameters, while LLM initialization biases early exploration.The two mechanisms are factored rewards and an LLM-driven warm start that prunes many suboptimal combinations.
  • F. Extended Discussion on MAB Optimization Efficiency: T=4 iterations are sufficient for the MAB to converge rapidly on an effective creative configuration and outperform Random Search.The result provides global search benefits without exhaustive sampling overhead.
  • G.1. Horizontal Directing through Hierarchical Parameterization: The Orchestrator synthesizes creative parameters into a top-down direction, decomposes it into technical directives, and injects those directives dynamically into sub-agents.This horizontal directing mechanism creates a unified high-level vision rather than passing raw categorical arms as static strings.
  • G. Creative Direction Examples: One iteration combines Informational, Vignette, and Clarity/Energy choices for functional utility, whereas another combines Transformational, Analytical, and Cinematic Premium choices for a logical manifesto.The examples show corresponding shifts from vibrant suburban transit and fast cuts to Premium Noir, chiaroscuro lighting, and reflective cinematography.
  • G.2. Visual Realization via Aesthetic Archetypes: Four aesthetic archetypes produce distinct visual treatments—Pop, Movie, Gallery, and Raw—while preserving brand logos and structural features across narrative arcs.The archetypes respectively emphasize recognition, prestige, material essentialism, and verité intensity; the consistent assets demonstrate preserved brand integrity.

H. Local Optimization via Agentic Self-Refinement … J. Keyframe Evaluation Prompt

Co-Director uses agentic self-refinement to correct storyline hallucinations and visual continuity failures, then evaluates storylines and keyframe sequences with structured, actionable rubrics. These evaluations emphasize product and character fidelity, narrative cohesion, engagement, prompt adherence, and diagnosis of the stage responsible for errors.

  • H. Local Optimization via Agentic Self-Refinement: The agentic self-refinement loop checks intermediate artifacts for logical and visual integrity, correcting storyline hallucinations and enforcing character consistency.It combines feedback descent with local refinement to improve sequence-level coherence.
  • H.1. Storyline Self-Refinement: The storyline loop refines scene-by-scene scripts against the product brief because language models may hallucinate product categories or misinterpret functional attributes.The example’s initial storyline incorrectly makes a spherical computer chassis function as a motorized standing desk.
  • H.1. Storyline Self-Refinement: The refined storyline restores the product’s intended role by depicting the AlphaFoo spherical computer chassis completing the high-speed 4K render.The correction moves the hero moment from the desk to the computer chassis.
  • H.2. Keyframe Self-Refinement: The keyframe loop addresses identity drift and environmental teleportation by jointly comparing generated frames rather than evaluating them in isolation.An MLLM visual auditor detects changes such as hair-color shifts or a studio changing from a garage to a corporate office, then reinjects corrections into the generation prompt.
  • I. Storyline Evaluation Prompt: The storyline evaluation prompt scores five dimensions from 0-20 points each and sums them into a total_score.The rubric covers hook quality, narrative arc and cohesion, product integration, engagement, and prompt adherence, with decisive, explanatory, actionable feedback.
  • J. Keyframe Evaluation Prompt: The keyframe evaluation prompt judges four generated images as one cohesive visual narrative using five 0-20-point criteria.It prioritizes character and product reference matching, internal consistency, narrative flow, product appeal and integration, emotional impact, and prompt adherence.
  • J. Keyframe Evaluation Prompt: The keyframe evaluator assesses sequences holistically, identifies whether failures originate in the storyline or image generation, and issues direct commands for the next iteration.Its structured output includes breakdown scores, MAB efficacy scores and justifications, primary_fault, actionable_feedback, and a total score from 0-100.

K. Final Video Evaluation Prompt · L. GenAd-Bench Evaluation System Prompts

The final-video evaluator scores generated advertisements for execution quality and strategic efficacy using supplied constraints, references, creative choices, theory, and storyboard context. GenAd-Bench separately evaluates advertisements on visual asset fidelity, demographic alignment, marketing appeal, and prompt-agnostic visual quality.

  • K. Final Video Evaluation Prompt: The final-video critic receives the video and original prompt, then evaluates the advertisement against structured constraints, reference visuals, creative configuration, theoretical definitions, and storyboard.The evaluator is instructed to scrutinize immersion-breaking details rather than relying on initial visual appeal.
  • K. Final Video Evaluation Prompt: Execution quality is scored across five dimensions—coherence, visual quality, engagement, prompt adherence, and logical consistency—with each dimension rated from 0 to 20 and summed into total_score.The prompt defines coherence, visual quality, engagement, and prompt adherence as strict evaluation dimensions; the output breakdown includes logical consistency.
  • K. Final Video Evaluation Prompt: Strategic efficacy is scored from 0 to 100 for creative strategy, narrative mode, and aesthetic archetype, evaluating whether the selected creative direction effectively produced the video.The framework distinguishes evaluating the strategic choice itself from evaluating faithful prompt following.
  • K. Final Video Evaluation Prompt: Efficacy scores must align directionally with execution quality, so strategic failures correspond to lower efficacy while high absolute quality corresponds to higher efficacy.The critic is explicitly told that a well-made but boring strategic choice should receive low strategy efficacy and low engagement.
  • K. Final Video Evaluation Prompt: The evaluator assesses narrative transportation and aesthetic alignment, including whether analytical, vignette, or narrative-drama structures and visual-auditory choices support the brand and intended mood.Examples include argument-based analytical structures, cohesive vignette atmospheres, beginning-middle-end drama, and archetypes such as Cinematic Premium or Kinetic Grit.
  • K. Final Video Evaluation Prompt: After scoring, the critic must provide timestamped or frame-specific feedback, identify the primary fault as storyline, image, or video, and issue a direct corrective instruction.The response must be a single JSON object containing detailed scores, efficacy justifications, feedback, and fault attribution.
  • L. GenAd-Bench Evaluation System Prompts: GenAD-Bench evaluates videos using brand-logo and product-image references plus six-point demographic constraints covering brand, product, target gender, age, location, and interest.Each of the four evaluation dimensions receives a score from 0 to 100.
  • L. GenAd-Bench Evaluation System Prompts: Its four dimensions are Visual Asset Fidelity, Demographic Alignment, Marketing Appeal, and prompt-agnostic Visual Quality, with reasoning required before each integer score in valid JSON.The dimensions cover asset identity and presentation, audience resonance, persuasiveness and product showcasing, and cinematic broadcast viability.
Loading 2604.24842v1…