acceptodds
Withdrawn from ICLR 2027

ProAgent: High-Fidelity Product Video Generation via a Multi-Agent Framework

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent text-to-video models have made advertising video generation attractive as a way to cut production time and cost, yet direct prompting often yields videos that drift from product identity, contain hallucinations, distort printed text, and fail to convey a clear commercial expression. We present ProAgent, the first to address high-fidelity product video generation—a training-free multi-agent framework that turns a user prompt identifying a specific product into an advertisement-ready video end-to-end. ProAgent grounds this intent in retrieved product evidence, plans a shot-level blueprint and rewrites it into a generation prompt under explicit product constraints, and iteratively improves the result through a closed-loop critique-and-recovery process. To evaluate the quality of generated advertising videos, we further introduce PV-Bench (Product Video Benchmark), a multi-dimensional benchmark scoring videos across six product-centric dimensions. Experiments across 12 product categories spanning major e-commerce industries and multiple state-of-the-art video generators show that ProAgent consistently outperforms baselines on both PV-Bench and general metrics for video quality and semantic fidelity.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-Agent, video generation, video generation benchmark, training free

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_sg6qOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces ProAgent, a training-free multi-agent framework for high-fidelity product advertising video generation. The system consists of context grounding through web/image retrieval, structured shot planning and prompt rewriting, iterative visual critique and targeted recovery, and an experience-learning module that reuses lessons from previous generation trajectories. The paper also introduces PV-Bench, containing 120 real e-commerce products from 12 categories and evaluating generated videos along six product-centric dimensions. Experiments on three video generators show that ProAgent improves over Direct Prompting and Single Rewriting on PV-Bench, general video-quality metrics, semantic/reference consistency metrics, and human evaluation.

Critical strengths
  1. Product advertising generation places stricter requirements on product identity, visual details, logos/text, hallucination avoidance, and commercial presentation than generic video generation. Explicitly formulating these requirements is useful and practically relevant.

  2. The combination of external product grounding, structured shot planning, prompt rewriting, visual critique, targeted recovery, and experience reuse forms a reasonably complete end-to-end workflow. Its training-free nature and compatibility with black-box generators also make the framework practically deployable.

  3. PV-Bench contains 120 products from 12 e-commerce categories, and the main experiments cover three different video generators. The reported improvements are not restricted to a single generator, which provides some evidence of cross-model generality.

  4. In addition to PV-Bench, the authors report VBench-Quality, VisionReward, VQAScore, DINO similarity, human evaluation, component ablations, iteration analysis, and backbone ablations. This provides a broader picture than relying on a single automatic metric.

Critical weaknesses
  1. ProAgent can repeatedly generate and inspect candidates until a stopping criterion or maximum number of attempts is reached, while DP and SR are single-pass baselines. Therefore, the current results cannot cleanly establish that critique/recovery is responsible for the improvement rather than simply having multiple opportunities to sample a better video. A crucial baseline would be DP and SR with the same number of generation attempts, followed by best-of-K selection or an equivalent fixed-budget protocol.
  2. ProAgent constructs a knowledge base from web search and a relatively large pool of retrieved images, including textual specifications, selling points, visual attributes, and multiple views. In contrast, DP and SR are given only the top-1 retrieved reference image, even though the paper describes this as “equal grounding.” This makes it difficult to isolate the contribution of multi-agent planning and recovery from the benefit of simply providing substantially richer product evidence.
  3. The six-dimensional checklist is evaluated by Gemini-3.1-Pro, and PV-Score uses manually chosen weights of 0.25/0.25/0.20/0.10/0.10/0.10. Although a human evaluation of generated outputs is provided, I do not see a direct validation of the benchmark judge itself, such as agreement between human annotations and MLLM checklist decisions, per-dimension accuracy, or sensitivity analysis for the PV-Score weights. This is particularly important because PV-Bench is presented as an independent contribution.
  4. There is no matched-budget repeated-sampling baseline, generic self-refinement baseline, or competitive iterative/multi-agent video-generation baseline. As a result, the experiments establish that ProAgent outperforms two relatively simple pipelines, but do not convincingly demonstrate that its specific multi-agent structure is necessary or superior to simpler test-time scaling strategies.
  5. The paper states that experience distilled from previous trajectories is stored and reused in later generations, such that even the initial prompt can benefit from prior experience. However, it is unclear whether this experience pool is reset between benchmark examples, generators, or categories, and whether benchmark products can influence later benchmark products. Because removing Experience Learning measurably reduces performance, this detail could materially affect the reported results and their reproducibility.
Questions for authors

What is the exact maximum and average number of video-generation calls used by ProAgent in the main results, and can the authors provide results against DP/SR with exactly the same generation budget (e.g., best-of-K)? Can the authors also evaluate a baseline that receives the same full retrieved knowledge base—including multiple images and web-derived product information—as ProAgent, so that the contribution of planning/critique/recovery can be isolated? How was the Gemini-3.1-Pro judge validated against human annotations for the six PV-Bench dimensions, and how sensitive are the conclusions to the chosen PV-Score weights? Finally, is the Experience Learning pool reset across products, categories, and generators, or can experience accumulated from earlier benchmark examples influence later ones?

Reviewer_81TTOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper defines high-fidelity product video generation and proposes ProAgent, a training-free multi-agent pipeline that retrieves product evidence, plans shots, rewrites prompts, generates with black-box text-to-video models, critiques the result, and applies targeted recovery until acceptance. It also introduces PV-Bench, 120 real e-commerce products scored by an MLLM judge on 26 checklist items across six product-centric dimensions (F/H/T/P/C/A). The reported experiments compare ProAgent with Direct Prompting and Single Rewriting on three generators, together with general video metrics, human ratings and component/backbone ablations.

Critical strengths
  1. A reusable product-centric benchmark. PV-Bench defines 120 real products across 12 categories and 26 binary checklist items over six dimensions, with an independent judge and a reproducible PV-Score formula. This is a concrete asset for evaluating product fidelity in generated advertisements.
  2. Broad and consistent empirical evidence. The evaluation covers three generators, two baselines, four general metrics, human ratings, component and backbone ablations and an iteration sweep. The direction of the product-fidelity gains is consistent across generators, and DINO similarity provides model-free corroboration.
Critical weaknesses
  1. The "first" claims are contradicted by prior work that is not cited. The paper claims the first task and the first training-free multi-agent framework, but AgenticGen (arXiv:2609.09187) already performs agentic advertising video generation, and DreamActor-H1 (arXiv:2506.10568) and AnchorCrafter (arXiv:2411.17383) already generate high-fidelity product-promotion videos. SCMAPR (arXiv:2604.05489) and GENMAC cover multi-agent prompt refinement and iterative verification. The contribution must be repositioned as an incremental benchmark/system contribution and compared with these works.
  2. The central attribution is not supported by controlled comparisons. DP/SR receive only a top-1 reference image while ProAgent uses a multi-image pool; ProAgent also uses more test-time compute; the iteration budget K for the main results is not reported; and the tables contain no error bars or significance tests. Table 2 shows that most of the gain comes from retrieval plus a single rewrite. The paper therefore does not establish that the multi-agent critique/recovery design causes the improvement, and the system cannot be reproduced from the released material (no code, data, benchmark link or full production prompts).
Questions for authors
  1. What are the iteration budget K, the number of generations and the LLM/VLM call counts for the main results, and how does ProAgent compare with best-of-K sampling and a Self-Refine baseline under the same budget?
  2. How is the experience pool initialized and updated? Does it accumulate across the 120 evaluation products, and what results are obtained with a fixed pool or a per-product reset?
  3. Exactly how many reference images do DP, SR and ProAgent receive? Do the reported gains persist when all methods receive identical references?
  4. Can you report product-level confidence intervals, multiple judge/model runs, and human-judge agreement for the checklist, and give the SR results under the weaker backbones of Table 4?
Reviewer_PBMQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ProAgent is a training-free multi-agent pipeline for converting product-oriented user intent into advertising videos. It retrieves textual and visual product evidence, creates a structured shot plan, rewrites constrained prompts, generates candidates, and iteratively critiques and repairs failures while storing reusable experiences. The paper also introduces PV-Bench, containing 120 real products across 12 categories and six fidelity and commercial dimensions. Across three video generators, ProAgent outperforms direct prompting and single rewriting on an MLLM judge, general video metrics, semantic similarity, and a small human study.

Critical strengths
  • PV-Bench targets an important fidelity problem using diverse real products and multidimensional criteria.
  • Diagnosis-driven recovery is more targeted than generic scalar self-refinement and works consistently across three generators.
Critical weaknesses
  • The comparison does not isolate multi-agent coordination from additional information and computation. ProAgent uses web evidence, an image pool, multiple specialist calls, and up to three recovery iterations, whereas baselines are one-pass. Without compute-matched best-of-(K), iterative self-refinement, retrieval-only, or reference-selection baselines, the gains cannot be attributed to the framework.

  • Core operations are specified only as black-box LLM calls, and the promised appendix containing prompts and hyperparameters is absent. PV-Bench relies principally on one checklist judge; the five-rater human study reports no agreement, uncertainty, or judge-human correlation, and stochastic generations are not repeated. These omissions prevent reproduction and reliable effect estimation.

Questions for authors
  • Under identical retrieved evidence and equal video-generation budgets, how does ProAgent compare with best-of-(K) selection and a strong single-agent iterative-refinement baseline? Please also report latency, API calls, and computational cost.

  • How was the Gemini judge validated against human annotations? Please report per-dimension agreement, judge-human correlation, repeated-judge stability, sensitivity to alternative judges and PV-Score weights, and confidence intervals for the main results.

  • Was the experience pool reset before each product or accumulated across PV-Bench? If accumulated, what evaluation ordering or held-out split prevents earlier test products from influencing later results?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.