acceptodds
Under review as a conference paper at ICLR 2027

Temporal Geometry Consistency for Local Structural Drift Verification in Generated Videos

Abstract

We study local structural drift in generated videos using predicted geometry. We introduce Temporal-Geometry Consistency (TGC), a cohort-relative score that combines four structural and directional-temporal readouts of predicted normals and depth through within-generator ranks. The factors capture temporal directional consistency of the geometry and relation fields, spatial structure in semantic-geometry agreement, and normal-direction consistency from an independent video-depth estimate. On the method-held-out GeneVA 407 set, TGC improves discrimination over the Relation baseline by +0.0367 AUROC under equal-generator, cluster-aware paired inference, with a 95% confidence interval of [+0.0040, +0.0693]. Direct agreement-based normal/depth rerankers do not improve over Relation, while matched-encoding diagnostics show that their agreement-amplitude readout is more sensitive to the tested encodings than the selected structural and directional-temporal summaries. TGC remains cohort-relative: it needs generator identity and a reference cohort. Its appearance-control non-inferiority guardrail is unresolved, so we present TGC as a candidate ranking verifier for local drift, not a calibrated or deployable score.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
video generation evaluation, geometric consistency, temporal consistency, local structural drift, post-hoc video verification

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_6UYXOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies post-hoc ranking of generated videos according to local structural drift, such as local shape deformation, inconsistent surfaces, or short-lived geometric distortions. It proposes Temporal-Geometry Consistency (TGC), a cohort-relative score that combines four readouts based on predicted surface normals and depth: directional coherence, structural variation of a semantic–geometry agreement field, directional residual coherence, and adjacent-frame depth-derived normal coherence. On the 407-pair GeneVA evaluation, TGC improves over the Relation baseline by approximately 0.0367 AUROC. The paper also reports ablations, encoding diagnostics, per-generator analyses, bootstrap confidence intervals, and a sealed same-engine evaluation. However, on the pre-registered VideoPhy-2 evaluation, the pooled improvement over Relation is small and statistically unresolved. The results are positive for three unseen generators and negative for two, and the primary fresh union does not show a clear improvement. The paper therefore provides evidence for a useful GeneVA-style cohort-relative verifier, while cross-generator generalization remains uncertain.

Critical strengths
  1. The paper targets a meaningful evaluation problem. Local structural drift can be missed by global video-quality metrics, and a post-hoc ranking tool could help prioritize clips for human inspection.
  2. The evaluation includes useful controls and diagnostics. Table 4 reports generator-level results and leave-one-generator-out analyses; Table 5 controls for encoding; Table 7 reports fresh evaluation blocks and confidence intervals; and Table 10 compares against external geometry baselines on a sealed GeneVA block. These analyses make the empirical claims more transparent.
Critical weaknesses
  1. The evidence for cross-generator effectiveness is limited. On the main GeneVA evaluation, TGC improves over Relation by +0.0367 AUROC (Table 7). However, on VideoPhy-2, the gain is only +0.0036 with a 95% confidence interval of [-0.0565, +0.0643]. The primary fresh union is -0.0086 [-0.1270, +0.1028], and the three-block union is -0.0007 [-0.0808, +0.0752]. Table 8 shows positive gains on three generators, while Table 9 shows negative gains on two. These results are consistent with the authors’ stated limitation, but they restrict the contribution to a narrower, benchmark- or cohort-dependent result.
  2. The paper does not yet establish that TGC measures the target artifact itself. The ablations in Section 4 and Appendix D show that the proposed readouts affect ranking performance, including the depth-only factor and rank-normalized baselines. However, these results do not directly establish that high TGC scores correspond to independently verified local structural drift rather than correlated factors such as motion complexity, generator identity, or scene content. Independent artifact-level annotations or controlled synthetic deformations would strengthen the causal interpretation of the score
Questions for authors
  1. Why is the four-factor construction necessary? The depth-only factor outperforms TGC on GeneVA (0.7241 vs. 0.7177, Appendix D) and has larger reported gains on both fresh blocks (Table 7). What evidence shows that the other three factors add useful information rather than dilute the strongest signal?
  2. What makes temporal geometry inconsistency specific to structural drift? Legitimate motion, occlusion, viewpoint changes, and geometry-estimation errors can also produce inconsistency, while persistent structural defects may remain temporally coherent. How does TGC distinguish these cases?
  3. What justifies combining within-generator ranks? Rank normalization discards magnitude and does not account for factor reliability or redundancy. What methodological or empirical evidence supports the proposed aggregation rule?
  4. Where is the independently confirmed improvement? The VideoPhy-2 gain is only +0.0036 AUROC, and both fresh-union estimates are near zero or negative (Table 7). What evidence establishes a meaningful advantage over Relation and the simpler depth-only comparator?
Reviewer_fdHNOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors of this paper describe how to rank generated videos so that a human reviewer sees clips with local structural errors first, such as a hand gaining an extra finger or a face warping. Clips are ranked within one generator’s batch, not across generators. The authors first report a negative result: their earlier geometry-based rerankers did not beat their in-house baseline, “Relation,” on 407 GeneVA pairs. They then propose TGC, a training-free score that combines four signals measuring how consistent predicted geometry is over time:

ϕdir\phi_{dir}: how stable the normal/depth directions are over time.

ϕgrad\phi_{grad}: how much the semantic–geometry agreement map changes across space.

ϕrel\phi_{rel}: how stable Relation’s direction field is over time.

ϕdep\phi_{dep}: how consistent surface normals are between adjacent frames, using a separate video depth model.

The authors show that each signal is converted to a rank within its generator, and the four ranks are averaged, across the 407 GeneVA pairs, TGC improves AUROC from 0.681 to 0.718 (gain +0.037, 95% CI [+0.004, +0.069]). A controlled experiment shows that compression strongly changes agreement-based scores, but lossless input gives no ranking advantage, and on the new pre-registered test sets (VideoPhy-2, Physion, and a sealed GeneVA replication), the gain over Relation is not confirmed. The combined estimate is about zero, as the authors state, and the sealed GeneVA set, TGC beats two external methods, GeCo and VIGOR (0.687 vs. 0.502 and 0.491). TGC passes a pre-set robustness check under lighting and exposure changes.

Critical strengths
  • The paper fixes its method and statistical tests before seeing the test labels. It uses paired statistics that treat each generator equally, and it reports the new test sets even though they do not confirm the main result. It also states its limits openly, for example, that the four-part design does not show a clear advantage over its depth-only part. This makes the numbers trustworthy, as other work in this area could reuse the protocol, where papers usually report a single pooled number.

  • The authors score the same generated videos in lossless and compressed form, which cleanly separates the effect of compression from the effect of content. The experiment shows that scores based on agreement levels move a lot under compression, while scores based on spatial structure move much less. This is useful for anyone building geometry-based checks for generated video.

Critical weaknesses

The method contribution is not supported beyond one development-adjacent test set, as it is acknowledged by two authors by the two key limits:

  • The gain over Relation is not confirmed on new test sets (Section 4.7).
  • The four-part design is not shown to be better than ϕdep\phi_{dep} alone (Section 4.4).

1. The only significant gain is on the 407-pair set, with a lower bound of +0.004. The paper states TGC was developed on Wan and frozen before these labels were used. But this is the same set on which the authors’ earlier rerankers failed, and that failure motivated TGC’s design. So the set is held out from TGC’s formula, but not fully from the research process.

2. No new test set confirms it because this includes a 200-pair replication with the same generators (+0.0151, CI [−0.027, +0.058]). The three-set combined estimate is −0.0007. The ϕdep\phi_{dep} alone does as well or better. It scores 0.724 vs. TGC’s 0.718 on the 407-pair set. It is the only part that beats Relation on all six test pools. On both new test sets where it was compared, its gain over Relation is larger than TGC’s (+0.020 vs. +0.004, and +0.023 vs. +0.015). The ϕdep\phi_{dep} is a simple frame-to-frame normal check from an off-the-shelf depth model. The simplest reading is that it carries most of the signal, and the other three parts add little that can be shown. The abstract also presents the sealed-block result only against GeCo and VIGOR (0.687 vs. ~0.50). It does not mention that Relation scores 0.666 on the same block, or that TGC’s gain over Relation there is not significant. Section 4.7 reports this and it should be written in the abstract as well.

3. TGC cannot be run without Relation as the definition of the ''Relation'' is described only briefly (App. B.1), not in enough detail to rebuild. All four TGC factors are computed on the shared node grid and support map, which appears to be Relation’s node-drift mass (App. B.1). ϕrel\phi_{rel} and ϕgrad\phi_{grad} use Relation’s own fields directly. The package reproduces the tables from saved scores, but the paper does not say whether Relation will be released. Without it, others cannot apply TGC to new videos. GeCo is used outside its stated scope. GeCo’s own paper describes it as a metric for artifacts in static scenes, but the test videos contain moving people and objects. Its near-chance result (0.50) is therefore expected, and the paper does not discuss this. Relation itself scores 0.666 on the same block. Beating GeCo and VIGOR there mainly shows that these tools do not fit this task, not that TGC is strong. The claim that TGC “stands clearly ahead” of the external baselines should be qualified.

The writing of the paper is hard to read, i will point out some terms which were not stated clearly anywhere.

  • Relation, the main baseline, is only named in the main text, its description is in App. B.1, it should be explained clearly in the main text to follow it easily what it means and where it comes from.
  • Many terms are used without definition, such as “fresh-headroom,” “hero quotas,” “same-engine replication,” “ledger,” “receipt,” and “readout.”
  • The abstract contains about a dozen numbers with confidence intervals, so the main result is hard to find.
  • At the end of p. 2 (L107), one sentence says the appearance-control margin is “unresolved” and the next says it is “established.” (L108-109) This reads as a contradiction, because it is not clear that the first sentence refers to the earlier, smaller control sets.
  • Some sentences are unclear, for example “Generalization to unseen generators therefore stays open, alongside the frozen candidate, the diagnosis and the protocol” (p. 3, L113-114).
  • Formatting errors: a broken paragraph header in (section 3.2), mismatched quote marks in (section 3.4), a double period at the end of App. E, and inconsistent generator names (Pika-1/pika1, VC-2/vc2, CogVideoX/cogvideo).
  • Sora and Pika, two of the three main generators, are not cited. (Section. 2) calls Physion “action-centric” and groups it with generated-video benchmarks but the Physion is a benchmark for predicting how simulated physical scenes will evolve, so can the authors explain this ?
Questions for authors

1. GeCo is designed for static scenes but why is it a fair baseline on dynamic GeneVA videos, and how were its settings chosen? Please also give the per-generator AUROCs for GeCo and VIGOR, shown only in Figure 1 and Figure 3(c), as a table.

2. (Section 4.4) says the four-part design is not shown to be better than ϕdep\phi_{dep} alone but what is the paired gain of full TGC over ϕdep\phi_{dep} alone on the 407-pair set and on each new test set? If it is zero or negative, what is the case for keeping all four parts?

3. How were the “strict local drift” labels created from public annotations? VideoPhy-2 labels rate physical commonsense and prompt adherence, and GeneVA provides free-text descriptions and boxes. What rules mapped these to drift / no-drift? The Ethics statement says an author did the internal annotations. Was there a second annotator or an agreement check?

4. Before freezing TGC on Wan, how many design variants were tried? Did the earlier reranker results on the 407-pair set guide the choice of the four parts?

5. Relation’s AUROC is below 0.5 on several sets: Wan 0.36, LTX 0.42, the 79-clip set 0.38, and CamI2V about 0.30, which App. D calls “near chance.” Is Relation’s score direction consistent across sets? Does this affect how the gains on those sets should be read?

6. The Physion block was reduced to 12 pairs after a quota shortfall, but it keeps equal weight with the 170-pair VideoPhy-2 block in the primary combined test. What effect size could this combined test detect? Also, how was Physion used to build a generated-video block, given that it is a simulated-scene prediction benchmark?

Reviewer_bHL8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper ranks generated videos for human inspection of local structural defects. Global quality metrics often overlook brief deformations, while direct normal–depth agreement provides limited additional discrimination. Temporal-Geometry Consistency (TGC) instead measures structural variation and directional consistency over time. Four factors capture geometry-direction coherence, semantic–geometry agreement gradients, relation-field coherence, and depth-derived normal consistency. Within-generator rank aggregation reduces score-scale differences, with controlled evaluations examining compression sensitivity and appearance perturbations. TGC improves GeneVA AUROC from 0.6808 to 0.7177, but unseen-generator gains remain inconclusive, and its cohort-relative scores are not calibrated defect probabilities.

Critical strengths
  1. Evidence-driven motivation. The documented failure of agreement-based reranking provides a concrete rationale for examining structural variation and directional temporal consistency.
  2. Careful diagnostic analysis. Encoding controls, rank-normalized comparisons, depth-only evaluation, and appearance perturbations examine alternative explanations and clarify the proposed mechanism.
  3. Strong statistical discipline. Frozen evaluation rules, generator-aware paired inference, fresh confirmation sets, and numerical reproduction fixtures improve transparency and limit selective reporting.
Critical weaknesses
  1. Unproven ensemble necessity. The depth-derived factor alone achieves 0.7241 GeneVA AUROC, exceeding the ensemble’s 0.7177, leaving the need for four components and equal weighting unresolved.
  2. Inconclusive generalization. VideoPhy-2 gains are negligible overall and negative for two of five unseen generators, while neither fresh evaluation union establishes superiority over Relation.
  3. Incomplete operational analysis. Scores depend on generator identities and reference cohorts, but stability under changing cohort composition and operation without generator labels remain unvalidated.
Questions for authors

Please see the weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.