acceptodds
Under review as a conference paper at ICLR 2027

Verify What Matters: Cost-Aware Long Video Generation with GIVER

Abstract

Although recent agentic systems can generate long multi-shot videos from storyboards, physical inconsistencies can persist, such as fingers passing through a cup during lifting. Repairing these defects requires identifying failures and guiding regeneration without altering the requested content. To this end, we introduce GIVER, an agentic framework that embeds verification-and-repair loops into the generation process. Specifically, GIVER constructs a fixed Subject–Object Interaction Graph (SOIG), whose nodes encode storyboard-derived visual requirements and whose edges encode their logical prerequisites. Using this graph, GIVER verifies reusable assets and repairs them when needed, then runs two local loops: L1 checks the first frame’s initial state, while L2 checks required video behavior and visible artifacts, retaining the selected first frame. Failed-node evidence guides targeted revisions to generation fields, and each new candidate is rechecked against the same requirements. On ViMax-Bench and ST-Bench-10, GIVER leads the compared pipelines in fidelity and physical plausibility under automated and blinded human evaluation, with automated severe-artifact rates of 0.70% and 2.80%, respectively. Additionally, our ablation study shows improved quality with approximately 24–25% fewer video generation and editing operations than two flat-verification configurations.

Primary area:
generative models
Keywords:
long video generation, multi-shot video generation, structured verification, test-time refinement, agentic video generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UQU2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GIVER checks and repairs generated videos against a fixed set of storyboard requirements. It uses a dependency graph to connect failed checks with targeted revisions, with separate loops for the first frame and the video. Experiments compare it with existing video pipelines and flat verification, using both automated and human evaluation.

Critical strengths
  1. The useful idea is keeping the goal fixed while changing how it is realized. Otherwise, a repair system can make a video look better by quietly dropping the difficult action. Checking fidelity alongside physical plausibility addresses this problem directly.

  2. The reported improvements are supported by blinded human ratings on both benchmarks. This gives me more confidence that the system improves what viewers actually see, beyond satisfying its own verifier.

Critical weaknesses
  1. I am less convinced about what the dependency graph itself contributes. The flat-verification comparison is useful, but compares complete configurations. A fixed checklist with the same questions and specific feedback might already provide much of the benefit. The paper needs to show when a dependency changes the diagnosis or leads to a better repair.

  2. The evidence is mainly about individual shots. A collection of plausible shots can still tell an inconsistent story—for example, an object’s state can change across cuts. As the authors acknowledge, the current results establish local repair quality rather than full-story continuity. This limits the broader long-video conclusion.

Questions for authors

With the same atomic checks and repair feedback, what improvement comes specifically from the dependency edges? Concrete cases where the graph changes the repair decision would help explain the contribution.

Reviewer_MwMhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces GIVER (Graph-guided Iterative Verification and Evidence-linked Repair), an agentic framework designed to address physical inconsistencies (e.g., interpenetration, unexplained pop-ins) in long, multi-shot video generation. GIVER repairs video defects by identifying failures and guiding object reconstruction without altering the requested prompt’s context. GIVER constructs a fixed Subject-Object Interaction Graph (SOIG) from a storyboard derived visual requirement, an directed acyclic graph where nodes represent atomic visual requirements and directed edges represent encoded logical prerequisites. The method includes a two-staged asset verification-and-repair loops using first-frame (L1) and subsequent video frames (L2). Validation is performed through automated tests (quantitative and qualitative) combined with human evaluation. An ablation test is included with the test/validation harness. Evidence from failed nodes targets specific prompt fields for revision without rewriting the entire storyboard contract. Evaluated on ViMax-Bench and ST-Bench-10, GIVER demonstrates improved script fidelity and physical plausibility while reducing video generation/editing operations compared to flat verification pipelines. However, performance is not consistent and uncorrected errors in the first frame will propagate through the rest of the frames in the proposed GIVER verification-and-repair loop. The paper fails to establish the reliability of the proposed mechanism or explain why the gains were significantly better than existing methods. Therefore, any claims to novelty disappear due to lack of evidence.

Critical strengths

There are a few strengths to this paper, including the attempt to address failure modes in generative AI in videos which is a challenging task. The staged workflow makes sense for a complex method such as the one discussed, and if properly elaborated on, would have strengthened the novelty which the manuscript was trying to solve. The contrast and comparison of the automated vice human evaluation is notable in that it attempts to calibrate a complex automated evaluation workflow. Finally, it is commendable that the authors recognized the weaknesses in their results, as it pertains to correlations. However, the reviewers would have had to go to the appendix to encounter that statement.

Critical weaknesses

The paper has multiple contradictions that fail to strengthen the argument. For example, the authors asserted that observations can be traced to evaluated media (page 5, paragraph 1), yet also assert that “category grouping branch in Figure 2 [does not] imply a prerequisite” (pg 4, paragraph 2). The verbiage in the first three sections on fixed graph of visual requirements is confusing at best.

The discussion on the test harness lacked details (and the appendix never elaborated on it either). Primarily, a description of ideal image conditions were never made, hence it was difficult to accept the assertions on performance. Similarly, Figures 1 and specifically Figure 3, show multiple “pipelines” where videos are not exactly the same for each “pipeline” test, biasing likely complex scenes unfairly to other methods, and likely introducing better scores for GIVER. If this was a deficiency on the LLM aspect of the method, then it was not discussed. A fair comparison with fair scoring would have attempted to ensure similarities in frames generated from prompts as possible. There were no assertions that the same prompt (prose, stance, etc) was used across all pipelines, which needs to be clarified.

The initial conditions should have been normalized across all tests. Therefore, the evidence is a bit lacking: For example, acceptance is underspecified. §3 defines a graph score but acceptance (as logic would dictate) depends on critical features, that then dictate the nature of failures, vetoes, applicability rules, and waivers. The discussion on waivers was short and could have used more elaboration. The repair methods were also mostly assertions lacking any implementation discussions. There were discussions on “requirements” which suggests that there may be bias on input prompts that can cascade through the workflow, never mind the fact that the initial images also likely can drive cascading effects into the workflow that would penalize one method over another.

There are a number of concerns regarding multiple aspects of GIVER. The graph score could have benefitted from more elaboration on the method and quantification. The repair mechanism is largely asserted, but bereft of information on how the repair is actually done, or what specific features trigger it. So repeatability may be difficult to perform. This is somewhat demonstrated by the discrepancies in the pipeline (results) illustrated in Figure 3, where a cascade of prompt inputs may have altered the complexity of initial scenes per pipeline, which in turn may have unfairly penalized certain pipelines when compared to GIVER. Finally, the verification and validation discussion (§ 4.2 and beyond) appear to have many nuances in scoring discrepancies. Discussion on the discrepancy in human evaluation versus automated methods is lacking. Some of the SF and Phys tests differed by more than 0.02, and others 0.15 respectively. The parameters for each test are also concerning, with ViMax-Bench having significantly more test samples than ST-Bench-10, likely biasing the results. The human test is also not in parity with the automated test, where human evaluations were consistent for both ViMax-Bench and ST-Bench-10 in terms of sample size compared to each other, but significantly less in sample for the automated test. Hence, it’s difficult to trust the results, as good as the assertions may be. It is not clear why GIVER had two tests (Figure 4) when the other pipelines/methods did not. There were no justifications given for this “retry” (§ 4.4, and Figure 4).

Therefore, this reviewer thinks there’s insufficient specifications or elaborations on assertions to strengthen claims made regarding performance. The evaluated outputs received favorable scores for the introduced method, but they fail to establish the reliability of the proposed mechanism or explain why the gains were significantly better than existing methods. Therefore, any claims to novelty disappear due to lack of evidence.

Questions for authors

The appearance of irreproducible results lowers the proposed novelty of the effort. Although the automated workflow appear to have promising performance, there are flaws in the validation despite attempts to mitigate it with standardized metrics. The inconsistencies in initial data conditions or parameters, and lack of discussion on (normalized) prompt designs/workflows leaves more questions than answers on the likely causes of behavioral deviations between each pipeline/method tested against GIVER. The acceptance criterion for the tests needs more discussion as does assessing behavior between the different methods to illustrate the strengths of GIVER. The dependency-aware/cost linking and assessment also fails to provide empirical support for its benefit. The paper makes a cost-benefit assertion but lacks discussion or quantification of improvements gained on the proposed novelty of GIVER. Finally, assertions on improvements to long videos fails to provide sufficient evidence. Combined with the failure to set initial frame conditions, the overall assertions cannot be sufficiently substantiated. Some specific suggestions are listed below. Clarify Repair Mechanics: Can you provide the exact refiner prompt template and rule set that converts node failure evidence into field edits?

Equal-Budget Baseline Comparison: Why were competing baselines not given an equal retry/refinement budget in Figure 4? What are the baseline scores if they are allowed an equivalent kk-attempt resampling strategy? In Figure 4 and § 4.4, GIVER is evaluated under both an "Initial" and a "Final" (retried) condition on a selected subset of 146 multi-attempt shots. However, none of the baseline pipelines (ViMax, MovieAgent, AniMaker) were granted an equivalent retry attempt or compute budget on these shots.

Address Evaluation Discrepancies: What explains the large discrepancy between human and MLLM scores (e.g., >0.4 gap on Phys)? How do you ensure that prompt modifications across iterations do not alter scene complexity relative to baselines (Fig. 3)? The authors provide no methodological justification for why only GIVER is permitted to retry and update its score while competing systems remain frozen at single-shot initial outputs. This creates a fundamental evaluation bias that invalidates the claimed performance gains in Figure 4.

Divergence Between Automated and Human Evaluation: The paper fails to discuss or justify the noticeable numerical discrepancies between MLLM automated metrics (ShotEval-MLLM) and human study results. On ST-Bench-10, Script Fidelity (SF) jumps by +0.134 between human ratings (0.657) and automated evaluation (0.791). For ViMax-Bench, physical plausibility (Phys) differs by >0.41 points (3.599 human vs. 4.010 automated).

Sample Size Imbalance (ViMax-Bench vs. ST-Bench-10): The automated benchmark evaluates 35 cases for ViMax-Bench but only 10 cases for ST-Bench-10, creating an asymmetric testing harness that risks biasing reported aggregate means. Human-to-Automated Parity Breakdown: While automated evaluations are heavily skewed in scale between benchmarks (35 vs. 10 cases), the human evaluation sample sizes are constructed symmetrically (N=768N=768 ratings for ViMax-Bench vs. N=2,322N=2,322 valid ratings for ST-Bench-10). This mismatch makes it difficult to align automated conclusions with human perception or establish metric reliability.

Prompt/Scene Cascading Differences: As illustrated in the qualitative comparisons (Figure 3), the input initial scenes appear to vary significantly in structural complexity across different pipelines. For instance, in panel (a), AniMaker and ViMax attempt complex geometric layouts (warped workpieces and block geometries) that are structurally distinct from the initial state evaluated by GIVER.

Unfair Penalty: If different pipelines undergo a cascade of prompt alterations that inherently change initial scene complexity, baseline systems are unfairly penalized. It is unclear whether GIVER performs better because of its graph mechanism or because baseline pipelines were evaluated on visually more intricate or drifting prompts.

Reviewer_kaMzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GIVER, an agentic pipeline for multi-shot video generation that wraps verification-repair loops around black-box image and video generators. From a storyboard it constructs a dependency-aware Subject-Object Interaction Graph whose nodes stay fixed across repair attempts; when a node fails, the associated visual evidence is routed to the stage that owns the corresponding generation field. Two loops follow: L1 validates and repairs the first frame's initial state, and L2 refines the video while holding the selected first frame fixed. Evaluation covers ViMax-Bench and ST-Bench-10, using automatic metrics and a blinded human study. In the shared-first-frame ablation, GIVER reports higher quality than the flat-verification configurations while using about 25% fewer video generation and edit operations. Appendices document the run protocol and the complete evaluator prompts.

Critical strengths
  1. The paper proposes a dependency‑aware fixed SOIG graph. Instead of relying on vague global scores, it uses node‑level visual failure evidence to conduct targeted repair, avoiding undesired modification to original storyboard requirements.
  2. The staged pipeline with separate L1 first‑frame loop and L2 video loop is well‑designed. Fixing the validated first frame in L2 avoids repeated image generation and makes video‑stage repair focus on temporal artifacts.
  3. Evaluation includes automatic metrics and blinded human studies. The fixed-first-frame ablation reports quality improvements with fewer video generation and editing operations.
  4. The appendices provide detailed implementation protocols and evaluator prompts, which support reproducibility. Appendices provide full implementation protocols and complete MLLM evaluator prompts, which is helpful for follow‑up research.
Critical weaknesses
  1. Since L2 keeps the first frame fixed, defects accepted in the L1 stage cannot be corrected afterwards. The paper mentions this limitation but lacks quantitative statistics showing how frequently this issue harms final video quality.
  2. The reported 25% reduction only counts video generation and editing operations. It ignores heavy overhead brought by graph compilation and repeated MLLM verification calls, so readers may overestimate the overall cost savings.
  3. The whole repair mechanism heavily depends on MLLM‑based node‑level judgments. There is no manual calibration for the verifier’s false‑positive and false‑negative rates, and baseline adaptation details are not sufficiently clear for fair comparison.
Questions for authors
  1. Since L2 cannot fix defects inherited from L1-accepted first frames, could the authors provide quantitative statistics to show how often such inherited flaws degrade final video quality?
  2. Could the authors add explicit discussion about the total computational overhead including graph compilation and repeated MLLM verification, to avoid over‑interpretation of overall cost benefits?
  3. Could the authors report the MLLM verifier's false‑positive / negative rates and clarify preserved / disabled modules for adapted baselines?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.