acceptodds
Under review as a conference paper at ICLR 2027

GeoRoute: Rerouting Video Diffusion Trajectories Via Single-State Geometry Injection

11 views

Abstract

Recent video diffusion models generate visually compelling videos, yet often fail to preserve coherent object geometry over time. Existing methods alleviate this problem by persistently incorporating 3D priors throughout denoising. However, as the video representation evolves, its compatibility with the 3D prior may vary substantially across denoising states. Persistent 3D injection may therefore introduce geometric information at poorly matched states, suggesting that 3D guidance could be applied more selectively. Motivated by this, we trace 3D-video compatibility along the denoising trajectory and find that higher compatibility is concentrated within a compact range of intermediate states, while the best-matched state varies across samples. We further find that, even at the best-matched state, different frequency components show different levels of compatibility. Together, these observations suggest that both when to inject the 3D prior and what 3D information to transfer should be determined adaptively. Based on these findings, we propose GeoRoute, a plug-in framework for adaptive 3D injection. Specifically, a Timestep Router selects a sample-specific compatible state, while a Spectral Router determines which frequency components to transfer. We further introduce a decoupled training strategy that separates 3D-video alignment from post-intervention denoising for stable optimization. In addition, we present GeoVideoBench, a reference-free benchmark for evaluating temporal geometric consistency. Extensive experiments across multiple video diffusion models show that GeoRoute consistently improves geometric consistency while preserving overall video quality. Code and website are available in the supplementary material.

Primary area:
generative models
Keywords:
Video Diffusion, 3D Priors, Adaptive Geometry Injection, Reference-Free Video Evaluation

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_jX9XOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GeoRoute, a plug-in framework for improving temporal geometric consistency in pretrained video diffusion models through selective 3D-prior injection. The key observation is that 3D–video latent compatibility varies along the denoising trajectory and across frequency bands, motivating adaptive decisions about both when and what geometric information should be injected. GeoRoute therefore combines a sample-specific Timestep Router with a Spectral Router and a decoupled training strategy to redirect the denoising trajectory using a text-generated 3D prior. The paper also introduces GeoVideoBench, a reference-free benchmark measuring local rigidity, long-range 3D consistency, global shape stability, and topology consistency. Experiments on Wan, HunyuanVideo, and CogVideoX show consistent improvements over the corresponding base models, while comparisons against GeoVideo and VideoGPA suggest smaller gains over existing geometry-aware approaches.

Critical strengths
  1. The adaptive geometry-injection formulation is intuitive and reasonably well motivated. The paper provides empirical evidence that 3D–video compatibility is concentrated within an intermediate denoising window, varies across samples, and changes across frequency bands. This directly motivates the Timestep Router and Spectral Router instead of introducing them purely as architectural heuristics.

  2. The method is evaluated across multiple video backbones and supported by extensive ablations. GeoRoute improves all three tested backbones, and the ablation study separately examines fixed versus adaptive routing, spectral routing, incorrect or blank priors, and persistent versus transient injection. These experiments provide useful evidence that the gains are associated with correctly matched 3D information and adaptive routing.

Critical weaknesses
  1. The improvement over the strongest geometry-aware baseline is quite small. On the shared CogVideoX backbone, GeoRoute only improves the VBench average from 96.42 for VideoGPA to 96.69 and GeoVideoBench Overall from 65.43 to 65.78. Several individual geometric metrics also differ by less than one point. Given this small margin, stronger independent evaluation is needed to establish that the improvement over existing geometry-aware methods is practically meaningful. Moreover, all four GeoVideoBench components differ from VideoGPA by less than one point (RSPS +0.63, L3RC +0.81, GSSC +0.40, PTC +0.05), and results appear to come from a single random seed. Without these, it is not possible to tell whether the reported margins exceed sampling noise.

  2. Performance on genuinely dynamic scenes is not sufficiently characterized. The paper reports aggregate geometric-consistency metrics and provides several dynamic qualitative examples, but it does not separately evaluate static scenes, camera-motion-dominated scenes, articulated subjects, and complex human–object interactions. This distinction is important because the injected TRELLIS prior is a prompt-level 3D representation, and the difficulty of maintaining geometry can differ substantially between rigid objects and genuinely dynamic scenes.

  3. The evaluation and practical-cost story is incomplete. GeoVideoBench is introduced by the same paper and achieves moderate human alignment, so the small SOTA differences would benefit from stronger independent evaluation. Moreover, the human preference study only compares GeoRoute with the corresponding base backbones, not directly against GeoVideo or VideoGPA. GeoRoute also introduces an 8–20% inference overhead and requires online TRELLIS generation of a 3D prior for every prompt. The evidence for GeoVideoBench is also weaker than presented. Its overall human correlation (ρ = 0.275) is on par with GPT-4o (0.287) and SSIM-sim (0.269), so it is not shown to be better aligned with human judgment than generic baselines, and both the alignment study and the user study rely on eight participants. GeoVideoBench also scores only the largest foreground object (a SAM 2 mask on monocular relative depth), which matches the object-level TRELLIS prior GeoRoute injects; background and scene-level geometry are not evaluated. Together with the absence of VBench Dynamic Degree, the current metrics cannot distinguish improved geometry from reduced motion.

Questions for authors
  1. Could the authors provide a quantitative breakdown for static versus genuinely dynamic scenes? In particular, I would like to see separate results for static objects/camera motion, articulated human or animal motion, and human–object or multi-object interactions. Since VideoGPA has shown that geometry-aware training can also improve dynamic scenes, it would be important to verify that GeoRoute’s gains are not dominated by rigid or relatively static cases.

  2. Could the authors provide stronger independent evaluation of the small gains over VideoGPA and GeoVideo? Since GeoRoute improves the strongest CogVideoX-based baseline by only a small margin, I would find direct human preference comparisons against VideoGPA/GeoVideo particularly useful, not only against the original backbone. An additional external human-aligned evaluator, such as VideoReward (Improving Video Generation with Human Feedback). And addtional 3D consistent metrics such as 3DCS, Epipolar Error etc.

6 comments
AuthorsOpenReview ↗

jX9X — Response 1/4

Thank you for the careful review and for recognizing the empirical basis of our two routers, the improvements across Wan, HunyuanVideo, and CogVideoX, and the controlled ablations on routing, prior matching, and injection duration. We have added multi-seed comparisons, independent evaluations, and quantitative breakdowns for dynamic scenes to address your questions.

Q1. Critical Weakness 1

The improvement over the strongest geometry-aware baseline is quite small. On the shared CogVideoX backbone, GeoRoute only improves the VBench average from 96.42 for VideoGPA to 96.69 and GeoVideoBench Overall from 65.43 to 65.78. Several individual geometric metrics also differ by less than one point. Given this small margin, stronger independent evaluation is needed to establish that the improvement over existing geometry-aware methods is practically meaningful. Moreover, all four GeoVideoBench components differ from VideoGPA by less than one point (RSPS +0.63, L3RC +0.81, GSSC +0.40, PTC +0.05), and results appear to come from a single random seed. Without these, it is not possible to tell whether the reported margins exceed sampling noise.

Multiple seeds and significance.

GeoRoute achieves higher VB and RC than VideoGPA in all five additional matched generation runs, and the improvements in VB and Motion Smoothness remain significant after Holm correction. The original single-seed results did not quantify variability across runs. We compared both methods on the shared CogVideoX backbone using five new generation seeds, calculated the aggregates separately for each seed, and report sample standard deviations, paired 95% confidence intervals, and two-sided paired t-tests. Holm correction covers all nine scores.

R1. Five-seed comparison. Δ denotes GeoRoute minus VideoGPA.

StatisticSubject consistency (SC)Background consistency (BC)Motion smoothness (MS)VBench Avg.RSPSL3RCGSSCPTCOverall RC
VideoGPA,mean±SD95.412±0.45595.594±0.40997.454±0.30096.153±0.15642.358±0.36675.224±0.64084.494±0.74051.876±1.25465.215±0.624
GeoRoute,mean±SD96.322±0.28995.790±0.12098.394±0.17696.835±0.14143.012±0.77176.172±0.10285.512±0.46452.194±0.66065.884±0.284
95% CI for Δ[0.057,1.763][-0.209,0.601][0.487,1.393][0.347,1.017][-0.473,1.781][0.126,1.770][0.049,1.987][-0.488,1.124][0.047,1.291]
Unadjusted p0.04150.25010.00450.00480.18230.03290.04330.33480.0405
Holm-adjusted p0.24280.54690.04040.04040.54690.23030.24280.54690.2428

The mean improvements are 0.682 in VB and 0.669 in RC, and every component metric has a higher mean. VB and Motion Smoothness have Holm-adjusted p values of 0.0404. For RC, the unadjusted confidence interval is positive, while the Holm-adjusted p value is 0.2428. These additional runs support a consistent observed advantage over VideoGPA, with stronger statistical evidence for VB than for RC.

Independent evaluation.

The added independent geometric metrics and VideoReward comparison favor GeoRoute over VideoGPA on the shared CogVideoX backbone. We used the requested 3DCS and Epipolar Error metrics and the external VideoReward evaluator to compare the two methods directly.

R2. External evaluation against VideoGPA on CogVideoX. Win rates are automated VideoReward preferences for GeoRoute over VideoGPA.

Method3DCS↓Epipolar Error↓Visual-quality win rate (VQ)↑Motion-quality win rate (MQ)↑Text-alignment win rate (TA)↑
VideoGPA0.5560.567---
GeoRoute0.5190.52356.67%64.33%66.00%

GeoRoute reduces both geometric errors and obtains VideoReward win rates above 50% in visual quality, motion quality, and text alignment. These results provide evidence of improved geometry and perceptual quality beyond GeoVideoBench. The new direct comparison is against VideoGPA; the comparison with GeoVideo is reported in the original main results. Our human preference study compares GeoRoute with its base backbones, so R2 adds an automated external preference assessment rather than a new human study against VideoGPA or GeoVideo.

We also evaluated Wan and Wan + GeoRoute with PDI-Bench as a separate geometry benchmark.

R3. PDI-Bench evaluation of Wan and Wan + GeoRoute.

MethodPDIS↓Scale↓Traj↓Rigid↓MathPass↑
Wan0.72631.32750.26170.576057.1%
+GeoRoute0.56880.92740.21690.530870.6%

All five reported PDI-Bench metrics improve with GeoRoute, including a PDIS reduction from 0.7263 to 0.5688. Together, these evaluations supplement the original benchmark with independent evidence on both the base-model improvement and the comparison with VideoGPA.

AuthorsOpenReview ↗

jX9X — Response 2/4

Q2. Critical Weakness 2

Performance on genuinely dynamic scenes is not sufficiently characterized. The paper reports aggregate geometric-consistency metrics and provides several dynamic qualitative examples, but it does not separately evaluate static scenes, camera-motion-dominated scenes, articulated subjects, and complex human–object interactions. This distinction is important because the injected TRELLIS prior is a prompt-level 3D representation, and the difficulty of maintaining geometry can differ substantially between rigid objects and genuinely dynamic scenes.

GeoRoute improves VB and RC in all three categories, including articulated motion and human–object or multi-object interactions. We compared Wan and Wan + GeoRoute on the same examples in three OpenVid subsets of 100 each, selected with ChatGPT assistance. Static covers static objects or camera motion, Articulated covers human or animal motion, and Interaction covers human–object or multi-object interactions.

R4. Results by scene category.

MethodSceneSCBCMSVBRSPSL3RCGSSCPTCRC
WanStatic94.1796.2998.6496.3744.3776.8787.3851.2066.19
+GeoRouteStatic98.4398.6298.9698.6754.2485.4789.0450.9669.03
WanArticulated93.0595.2198.4695.5742.8073.8985.3450.7264.72
+GeoRouteArticulated97.9797.8598.6798.1653.2983.1788.9652.0469.07
WanInteraction91.0292.5896.4493.3538.7669.3581.1745.6259.90
+GeoRouteInteraction97.4597.2898.6097.7853.1383.0288.7651.8968.90

The interaction subset has the largest improvement in both aggregates. These results show that the gains extend to the tested dynamic scenes. The geometric scores follow the principal-foreground-object protocol described in Q3.

Q3. Critical Weakness 3

The evaluation and practical-cost story is incomplete. GeoVideoBench is introduced by the same paper and achieves moderate human alignment, so the small SOTA differences would benefit from stronger independent evaluation. Moreover, the human preference study only compares GeoRoute with the corresponding base backbones, not directly against GeoVideo or VideoGPA. GeoRoute also introduces an 8–20% inference overhead and requires online TRELLIS generation of a 3D prior for every prompt. The evidence for GeoVideoBench is also weaker than presented. Its overall human correlation (ρ = 0.275) is on par with GPT-4o (0.287) and SSIM-sim (0.269), so it is not shown to be better aligned with human judgment than generic baselines, and both the alignment study and the user study rely on eight participants. GeoVideoBench also scores only the largest foreground object (a SAM 2 mask on monocular relative depth), which matches the object-level TRELLIS prior GeoRoute injects; background and scene-level geometry are not evaluated. Together with the absence of VBench Dynamic Degree, the current metrics cannot distinguish improved geometry from reduced motion.

Independent evaluation and direct human comparisons.

The external comparisons in Q1 provide evidence beyond GeoVideoBench. Our human preference study compares GeoRoute with its base backbones; the added comparison with VideoGPA uses automated VideoReward preferences, not a new human study against VideoGPA or GeoVideo.

Online prior generation and inference cost.

Including online TRELLIS generation for each prompt, GeoRoute adds 9.32%–23.23% end-to-end inference time, mainly for prior generation, compatibility evaluation, and prior-feature processing. The original denoising category included 3D-to-2D projection, tensor concatenation, and sparse/dense point-cloud conversion; R5 reports these operations separately under routing/prior processing.

R5. End-to-end inference latency in seconds.

MethodTRELLISRouting/priorDenoisingOthersTotalOverhead
Wan2.1--203.916.08209.99-
Wan2.1 + GeoRoute7.0832.74204.175.89249.8819.00%
CogVideoX--211.3220.78232.10-
CogVideoX + GeoRoute9.0443.33212.9620.69286.0223.23%
HunyuanVideo--1174.8138.341213.15-
HunyuanVideo + GeoRoute17.1995.181175.2238.581326.179.32%

Denoising itself increases by only 0.26–1.64 seconds across the three backbones. A single injection still requires constructing the geometric representation and evaluating compatibility before transferring it, which explains the additional processing cost.

AuthorsOpenReview ↗

jX9X — Response 3/4

Q3 continued — benchmark evidence and motion.

Benchmark scope, human alignment, and the eight-rater study.

GeoVideoBench evaluates local rigidity, long-range 3D consistency, global shape stability, and silhouette topology of the principal foreground object; its validation combines controlled geometric tests with moderate agreement with human judgments.

For multi-object scenes, PTC can capture topology changes visible in the retained foreground silhouette, including changes in holes, skeleton branches, and thin structures. The correctness of interactions themselves, including semantics and physical dynamics, is outside this scope. The protocol also does not assess full background or scene-level 3D geometry. VBench Background Consistency separately measures temporal background consistency, and the external evaluations in R2–R3 provide additional evidence beyond GeoVideoBench.

The controlled validation in original Appendix C.4 uses 300 rendered videos with known geometry from six articulated assets and two camera paths. Corruption severity correlates with RC at −0.672 using oracle geometry and −0.708 through the RGB pipeline. Original Tables 7–9 test sensitivity to targeted geometric corruption and nuisance changes in lighting, appearance, and camera motion. Reliability weights are fixed across methods, and original Table 10 reports gains under both equal weighting and RC.

Within our human-alignment study, RC has Spearman ρ = 0.275 and pairwise agreement of 66.0%. The cited GPT-4o (0.287) and SSIM-sim (0.269) correlations provide context from different tasks and protocols; they do not establish a ranking under a common evaluation. To make agreement among our raters explicit, we additionally calculated Krippendorff's alpha for the blinded preference study.

R6. Krippendorff's alpha with 95% confidence intervals.

CriterionAlpha95% CI
Visual quality0.3842[0.3227,0.4410]
Geometric consistency0.3385[0.2157,0.4532]

Both estimates indicate modest inter-rater agreement. The preference and metric-alignment blocks use the same eight participants; the 16,534 criterion-specific preference judgments include ties and repeated assessments. Alpha measures agreement among raters, while metric–human correlation measures alignment between metric scores and human judgments.

Motion presence and reduced-motion controls.

GeoRoute maintains Dynamic Degree on all three backbones and improves VB and RC in every baseline motion group, including the high-motion group. We first report VBench Dynamic Degree, which measures whether motion is present.

R7. VBench Dynamic Degree (%).

BackboneBase+GeoRoute
Wan56.2756.54
CogVideoX66.3766.52
HunyuanVideo59.3459.38

Dynamic Degree is slightly higher with GeoRoute on each backbone. To test different motion levels, we divided Wan's VBench outputs into three groups using the mean magnitude of the largest 5% of inter-frame pixel flows. Each paired GeoRoute output retains its Wan counterpart's group, so the comparisons use the same examples.

R8. Results by baseline motion magnitude.

MethodMotionSCBCMSVBRSPSL3RCGSSCPTCRC
WanLow93.9496.5298.3996.2844.7177.1287.0551.4466.30
+GeoRouteLow98.6898.3999.2198.7653.9885.7289.2850.7268.99
WanMid92.8395.4298.2295.4942.5673.6585.6150.9664.83
+GeoRouteMid98.1997.6498.4398.0953.0482.9388.7252.2969.06
WanHigh92.7793.2799.1695.0643.4471.5987.9250.4864.93
+GeoRouteHigh96.8997.6498.2597.6052.5881.2888.2852.6668.79

The high-motion group improves by 2.54 VB and 3.86 RC, and both aggregates also improve in the low- and medium-motion groups. Individual dimensions do not improve uniformly: high-motion MS decreases by 0.91 points and low-motion PTC by 0.72 points.

The reduced-motion controls explain why motion and geometric consistency should be reported together. We randomly sampled 800 videos generated on the VBench prompt set and evaluated each original alongside slowed-resampling and repeated-single-frame versions.

R9. Motion controls.

VersionDynamic Degree (DD, %)RSPSL3RCGSSCPTCRC
Original53.1353.2083.3188.7651.8968.95
Slowed11.3853.6784.2289.4353.2770.02
Repeated frame0.0057.2185.9390.0756.7872.53

Reduced motion can increase geometric-consistency scores; a repeated frame is geometrically stable but contains no motion. GeoRoute maintains Dynamic Degree (R7), and VideoReward provides complementary motion-quality and text-alignment evidence (R2). Dynamic Degree measures motion presence, not exact amplitude or action-specific correctness.

Reviewer_UdnwOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GeoRoute, a framework that improves geometric consistency in video generation by injecting a 3D prior at a single intermediate diffusion state. The same text prompt is provided to TRELLIS and the video generator, and the resulting 3D features are converted into a tri-plane representation. A Timestep Router selects an injection state within a candidate window based on compatibility between the 3D features and video latents. A Spectral Router predicts the frequency range and injection strength, modifies the video latent once, and allows denoising to continue. Training separates the latent-alignment objective for the 3D adapter from the modified denoising objective for the Spectral Router and video LoRA. The paper also introduces GeoVideoBench, which evaluates local rigidity, long-range 3D consistency, global shape, and silhouette topology, and validates GeoRoute on Wan, HunyuanVideo, and CogVideoX.

Critical strengths

The paper investigates compatibility between 3D priors and video representations across denoising states. Selecting the injection timestep, frequency range, and strength provides a coherent approach to controlling geometric intervention.

Table 4 includes LoRA under matched training settings, fixed and adaptive routing, incorrect and blank priors, and persistent injection throughout denoising or within the candidate window. These comparisons help distinguish the effects of additional fine-tuning, prior matching, and injection strategy.

GeoRoute improves all three backbones, with average gains of 5.84 points on GeoVideoBench and 1.52 points across the three reported VBench metrics. Blinded geometric-consistency preference rates range from 66.43% to 75.39% over the corresponding baselines, providing evidence beyond automatic metrics.

GeoVideoBench includes controlled geometric corruptions, nuisance perturbations, and human-alignment analysis. The appendix also describes gradient paths, training configurations, and inference overhead, helping readers understand the implementation and its limitations.

Critical weaknesses
  1. The evaluation does not sufficiently exclude reduced motion as an explanation for improved consistency. The reported VBench metrics measure subject consistency, background consistency, and motion smoothness, but not motion magnitude or action adherence. Moreover, global similarity and silhouette-stability criteria may penalize legitimate articulation or deformation. The relationship between geometric improvement and motion fidelity therefore remains insufficiently established.

  2. The independent benefit of sample-adaptive routing is not fully isolated. 3D foundation priors and low-frequency latent interventions have precedents in Orbital and FreeInit, respectively. The ablations support GeoRoute’s adaptive single-injection design, but lack a learned, globally shared spectral-parameter baseline and a direct analysis linking compatibility scores to final geometric quality.

  3. The role of static 3D priors in complex dynamic scenes remains unclear. A prompt-generated prior may differ from the video in object identity, pose, or layout. Its utility under these discrepancies and partial prior errors is not systematically evaluated. Incorrect-prior controls support the importance of matched conditioning, but do not fully separate semantic matching from geometric-structure matching.

  4. Metric robustness limits the interpretation of quantitative gains. In Table 9’s controlled experiments, RSPS responds more strongly to lighting and camera changes than to its targeted geometric corruption. Reliability weighting and human evaluation help, but independent geometric evaluation or tests with ground-truth geometry are still needed to substantiate the interpretation of the gains.

Questions for authors

Could the authors verify the improvements using an independent evaluation framework, such as the cited PDI-Bench, or controlled data with ground-truth geometry? How does GeoVideoBench distinguish geometric errors from lighting changes, camera motion, and legitimate articulation?

Could the authors report dynamic degree, action adherence, and results grouped by motion magnitude? Could repeated-frame or reduced-motion diagnostic controls help assess whether the metrics distinguish geometric improvement from reduced motion?

Could sample-adaptive spectral parameters be compared with equally trainable parameters shared across all samples? Across candidate injection states, how well does compatibility predict final geometric quality, and how far is the first detected local minimum from the injection state yielding the best final result?

Beyond incorrect-prior controls, could priors from the same category but with different shapes or poses help separate semantic and geometric contributions? For multi-object prompts, which objects does the prior represent, and how do partial errors in that prior affect generation?

12 comments
AuthorsOpenReview ↗

Udnw — Response 1/4

Thank you for the detailed review and for recognizing the compatibility analysis, the controlled ablations, the improvements across three backbones and in the blinded human study, and the validation of GeoVideoBench. Your questions led us to add evidence on motion, sample-adaptive routing, prior variation, and independent geometric evaluation.

Q1. Critical Weakness 1

The evaluation does not sufficiently exclude reduced motion as an explanation for improved consistency. The reported VBench metrics measure subject consistency, background consistency, and motion smoothness, but not motion magnitude or action adherence. Moreover, global similarity and silhouette-stability criteria may penalize legitimate articulation or deformation. The relationship between geometric improvement and motion fidelity therefore remains insufficiently established.

GeoRoute maintains Dynamic Degree on all three backbones and improves both aggregate scores in every baseline motion group, including high-motion videos. We additionally evaluate motion quality and text alignment with VideoReward and test the metrics with deliberately reduced-motion controls.

R1. VBench Dynamic Degree (%).

BackboneBase+GeoRoute
Wan56.2756.54
CogVideoX66.3766.52
HunyuanVideo59.3459.38

We group Wan outputs by baseline optical-flow magnitude and assign each paired GeoRoute output to the same group, keeping the compared examples matched.

R2. Results by baseline motion magnitude. Motion is the mean of the largest 5% of adjacent-frame pixel-flow magnitudes.

MethodMotionSubject consistency (SC)Background consistency (BC)Motion smoothness (MS)VBench Avg.RSPSL3RCGSSCPTCOverall RC
WanLow93.9496.5298.3996.2844.7177.1287.0551.4466.30
+GeoRouteLow98.6898.3999.2198.7653.9885.7289.2850.7268.99
WanMid92.8395.4298.2295.4942.5673.6585.6150.9664.83
+GeoRouteMid98.1997.6498.4398.0953.0482.9388.7252.2969.06
WanHigh92.7793.2799.1695.0643.4471.5987.9250.4864.93
+GeoRouteHigh96.8997.6498.2597.6052.5881.2888.2852.6668.79

High-motion VB and RC improve by 2.54 and 3.86. Individual dimensions show a trade-off: low-motion PTC decreases by 0.72, and high-motion Motion Smoothness by 0.91.

R3. Motion controls on 800 randomly sampled VBench generations.

VersionDynamic Degree (DD, %)RSPSL3RCGSSCPTCRC
Original53.1353.2083.3188.7651.8968.95
Slowed11.3853.6784.2289.4353.2770.02
Repeated frame0.0057.2185.9390.0756.7872.53

Reduced motion increases temporal geometric stability, so RC increases in R3. GeoRoute's Dynamic Degree does not decrease in R1, and the external VideoReward comparison gives motion-quality and text-alignment win rates of 64.33% and 66.00% against VideoGPA (R10). These are complementary checks of motion presence and overall motion quality; action-specific adherence is not separately quantified. The effects of legitimate articulation and nuisance changes are discussed in Q4.

AuthorsOpenReview ↗

Udnw — Response 2/4

Q2. Critical Weakness 2

The independent benefit of sample-adaptive routing is not fully isolated. 3D foundation priors and low-frequency latent interventions have precedents in Orbital and FreeInit, respectively. The ablations support GeoRoute’s adaptive single-injection design, but lack a learned, globally shared spectral-parameter baseline and a direct analysis linking compatibility scores to final geometric quality.

Compared with learned shared spectral parameters, sample-adaptive routing improves VB from 97.73 to 98.15 and RC from 68.63 to 68.95. Both configurations learn the cutoff, boundary sharpness, and injection strength; the shared control uses one set for all samples, while the router predicts them from compatibility and timestep information.

R4. Learned shared and sample-adaptive spectral parameters on Wan.

ParametersSCBCMSVBRSPSL3RCGSSCPTCRC
Shared97.5797.2698.3597.7352.0181.2387.9652.6468.63
Adaptive97.9297.8998.6398.1553.2083.3188.7651.8968.95

Sample adaptation improves three geometric dimensions and both aggregates, while the shared setting has higher PTC. Orbital and FreeInit provide precedents for 3D priors and low-frequency interventions; our contribution is compatibility-guided selection of a single injection state and adaptation of the transferred spectrum and strength.

Compatibility and final quality. We tested the selected state and offsets of two and five scheduler steps. Both aggregate scores decrease as the offset increases.

R5. Injection-state perturbation on VBench. Offset 0 is the router-selected state.

OffsetSCBCMSVBRSPSL3RCGSSCPTCRC
097.9297.8998.6398.1553.2083.3188.7651.8968.95
296.5796.4497.9696.9951.2682.3988.3751.6268.36
595.4395.8297.1996.1550.0682.1288.0451.2467.92

The selected state gives the highest RC in 48 of 50 videos and the highest VB in all 50 among these three candidates. In the inspected compatibility curves, the first local minimum also has the lowest compatibility distance.

R6. Per-video rankings among the three tested candidates.

StatisticVBenchGeoVideoBench
0-step best50/5048/50
Pairwise agreement with 0>2>5150/150148/150
Mean rho(offset,quality)-1.000-0.980
Mean best-candidate gap (steps)00.08
Maximum gap (steps)02

The two remaining RC rankings are 2>0>5. Pairwise agreement compares each video's three candidate pairs with 0>2>5. The reported rho measures the within-video association between offset magnitude and final quality, averaged over videos. The selected state is on average 0.08 steps from the best RC candidate, with a maximum gap of two. This is a local perturbation test; the correlation between raw compatibility scores and final quality over every state in the full window was not measured.

AuthorsOpenReview ↗

Udnw — Response 3/4

Q3. Critical Weakness 3

The role of static 3D priors in complex dynamic scenes remains unclear. A prompt-generated prior may differ from the video in object identity, pose, or layout. Its utility under these discrepancies and partial prior errors is not systematically evaluated. Incorrect-prior controls support the importance of matched conditioning, but do not fully separate semantic matching from geometric-structure matching.

GeoRoute gives stable scores across five priors sampled from the same prompt, with VB of 98.147±0.034 and RC of 68.977±0.081. We vary the 3D prior while holding the prompt and video-generation seed fixed.

R7. Five 3D seeds. SD is the sample standard deviation across runs.

3D seedSCBCMSVBRSPSL3RCGSSCPTCRC
198.0197.8898.6298.1753.1483.2388.9352.0069.04
297.9997.9098.5098.1353.1183.3088.6751.9668.95
397.8797.8198.6198.1053.0583.1788.7151.8368.87
498.0597.8798.6198.1853.3683.2788.6651.9068.94
597.9897.7098.8198.1653.1983.3588.8552.0869.08
Mean±SD97.980±0.06797.832±0.08198.630±0.11298.147±0.03453.170±0.11883.264±0.06888.764±0.12051.954±0.09568.977±0.081

Semantic and geometric matching. A shared prompt gives semantically related conditioning while allowing the sampled shape, pose, or layout to differ. R7 tests robustness to this natural prior variation; original Table 4's incorrect-prior control establishes the importance of a relevant prior. The seed experiment does not separately identify the contributions of shape, pose, and semantics, which would require independently controlled same-category variations.

Complex scenes and prior coverage. GeoRoute also improves both aggregates on the tested articulated and interaction scenes. We compare the same examples in three OpenVid subsets of 100 each, selected with ChatGPT assistance: static objects or camera motion, human or animal articulation, and human–object or multi-object interactions.

R8. Results by scene category.

MethodSceneSCBCMSVBRSPSL3RCGSSCPTCRC
WanStatic94.1796.2998.6496.3744.3776.8787.3851.2066.19
+GeoRouteStatic98.4398.6298.9698.6754.2485.4789.0450.9669.03
WanArticulated93.0595.2198.4695.5742.8073.8985.3450.7264.72
+GeoRouteArticulated97.9797.8598.6798.1653.2983.1788.9652.0469.07
WanInteraction91.0292.5896.4493.3538.7669.3581.1745.6259.90
+GeoRouteInteraction97.4597.2898.6097.7853.1383.0288.7651.8968.90

TRELLIS receives the full prompt, but the objects represented depend on its output and may omit background objects. Partial prior errors may transfer incorrect structure; their individual effects have not been isolated. The available evidence is stability across sampled priors in R7 and improved foreground geometry in R8.

GeoVideoBench scores the principal foreground object's geometry. PTC can reflect changes in holes, branches, or thin structures visible in the retained silhouette; interaction correctness itself is outside this scope. At inference, the trained adapter and routers use the sampled prior and current video latents, without the low-frequency training target.

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

This paper proposes GeoRoute, a plug-in framework that injects a generative 3D prior into a video diffusion model at a single, adaptively selected denoising state. The core contribution is the routing mechanism designed from two empirical observations: compatibility concentrates within a compact intermediate window while the best point varies across samples, and the band-wise correspondence changes with the denoising state; the Timestep Router and the Spectral Router decide when and what to inject, respectively, and a decoupled training strategy ensures stable optimization. A reference-free benchmark, GeoVideoBench, is further proposed, which evaluates four geometric failure dimensions with a reliability-calibrated aggregate score. Experiments on three backbones report consistent geometric gains together with user-study preferences.

Critical strengths
  1. The framework is logically well structured. Decomposing geometric intervention into two decisions, when to inject and what to inject, handled by two routers matches the claim of the title strictly; the two routers are not designed arbitrarily, as the authors provide measured compatibility observations along denoising trajectories as evidence, corresponding one-to-one to the two findings in the motivation.
  2. The ablation and controlled comparisons are fairly complete, and the core claims are directly verified rather than supported only by overall performance gains. Dedicated controls cover the effect of the additional fine-tuning on the same extra data, the necessity of the main components, the contribution of each router, and the superiority of single-state injection, and a prior-replacement experiment further indicates that the gains come from correctly matched geometric information.
  3. The construction and validation of the proposed benchmark are methodologically sound. Each metric is equipped with targeted geometric corruptions and non-geometric nuisance controls such as lighting, appearance, and camera-path changes, verifying sensitivity to the targeted corruption and quantifying how differently each metric responds to non-geometric nuisance factors, so that metrics whose nuisance response dominates are downweighted; the aggregation weights are reliability-calibrated, and the ranking conclusions are consistent under both the equal-weight and the reliability-calibrated aggregations; the accompanying human preference study is blinded with a complete protocol and covers both the method-preference and the metric-alignment blocks. This validation procedure goes beyond common practice for self-constructed benchmarks.
  4. The method demonstrates good plug-in generalizability. Consistent geometric improvements without regressions in overall video quality are obtained on three video diffusion backbones with different architectures; the inference overhead is separately measured and decomposed by component, with the extra cost mainly from 3D-prior generation and geometry-aware denoising. The practical positioning is quantitatively supported.
Critical weaknesses
  1. The 3D prior is paired with the video only at the prompt level, and the instance-level correspondence and applicability boundary deserve discussion. The prior is sampled directly from the text prompt; during training the adapter is distilled toward the low-frequency component of the same video, whereas this target does not exist at inference, so the prior may not match the instance geometry actually presented in the video, and the ablation indeed shows that a mismatched prior harms the results. Moreover, the benchmark scores only a single principal foreground object, and multi-object and interaction scenes are not systematically evaluated. The authors are encouraged to explicitly discuss these two boundaries.
  2. The Spectral Router contributes negatively on the topology dimension. In the ablation, the full method scores lower on topology than three simpler variants, namely fixed routing, adaptive timestep, and window injection, and on Wan the improvement on this dimension is also the weakest among the four geometric dimensions in the main results; the paper neither analyzes this counter-intuitive result nor mentions it in the limitations. One possible explanation is that spectral injection is low-pass filtering and can hardly convey the mid- and high-frequency information needed to repair topological errors, but topological corruption is not necessarily determined by high-frequency information alone, and the paper provides no correspondence between frequency bands and corruption types to verify this, so the explanation remains a hypothesis. In addition, the key hyperparameters governing routing and injection are given directly, and the ablation includes no sensitivity analysis; a supplementary analysis is suggested.
  3. Results are reported as single point estimates, and the paper reports no variance, multiple random seeds, or significance tests, so the narrow margins over the strongest competitor can hardly support a stable ranking. Several results lack mechanistic explanation: the subject-consistency gain on one backbone far exceeds that on the other two, and single-time injection nevertheless increases the denoising time, for which the paper offers an attribution but no explanation; the selection criterion of the qualitative samples is also unstated. Furthermore, the human-alignment argument juxtaposes the proposed metrics with numbers reported by other works on different tasks and provides no inter-rater agreement reference, making the absolute level of the scores hard to calibrate.
  4. The core novelty lies mainly in the idea of routing-based single-point injection, while the remaining components are combinations of existing techniques, such as a linear-probe-like compatibility measure, a low-pass mask with statistical matching, and standard LoRA adaptation. Compared with the most related injection-based predecessors, the essential differences reduce to the choice of the injection time and the single transient intervention, and the paper does not directly compare against these methods experimentally. The incremental contribution relative to prior work should be articulated more clearly.
Questions for authors

see above weakness

5 comments
AuthorsOpenReview ↗

GtiM — Response 1/4

Thank you for the careful review and for recognizing the empirical motivation of our routing design, the controlled ablations, the benchmark validation, and the generalizability of our method across video diffusion backbones. We address your comments below.

Q1. Critical Weakness 1

The 3D prior is paired with the video only at the prompt level, and the instance-level correspondence and applicability boundary deserve discussion. The prior is sampled directly from the text prompt; during training the adapter is distilled toward the low-frequency component of the same video, whereas this target does not exist at inference, so the prior may not match the instance geometry actually presented in the video, and the ablation indeed shows that a mismatched prior harms the results. Moreover, the benchmark scores only a single principal foreground object, and multi-object and interaction scenes are not systematically evaluated. The authors are encouraged to explicitly discuss these two boundaries.

Prompt-level pairing and inference.

The prior is better viewed as semantically related conditioning with coarse structural information, rather than a pixel-wise correspondence map to the video. We project the 3D feature volume onto three orthogonal planes by averaging and combine the resulting compact features; we do not assume a one-to-one correspondence with video pixels.

The low-frequency video target is used only for training. At inference, the trained adapter and routers use the sampled prior and current video latents.

Original Table 4 reports RC of 68.96 with the prompt-paired prior and 61.01 with an incorrect prior, showing the importance of using a relevant prior. To further test naturally occurring variation under a shared prompt, we added a multi-seed experiment with different 3D priors (R1): VB is 98.147±0.034 and RC is 68.977±0.081. GeoRoute remains stable across the natural variation in these five prompt-paired priors. The geometric quality and object coverage of the prior depend on the 3D generator.

R1. Results across 3D-prior seeds on Wan. The last row reports mean ± sample standard deviation.

3D seedSCBCMSVBRSPSL3RCGSSCPTCRC
198.0197.8898.6298.1753.1483.2388.9352.0069.04
297.9997.9098.5098.1353.1183.3088.6751.9668.95
397.8797.8198.6198.1053.0583.1788.7151.8368.87
498.0597.8798.6198.1853.3683.2788.6651.9068.94
597.9897.7098.8198.1653.1983.3588.8552.0869.08
Mean±SD97.980±0.06797.832±0.08198.630±0.11298.147±0.03453.170±0.11883.264±0.06888.764±0.12051.954±0.09568.977±0.081

Multi-object and interaction scenes.

GeoRoute's gains extend to the tested multi-object and interaction scenes, with higher VB and RC in all three scene categories below. We compared Wan and Wan + GeoRoute on three OpenVid subsets of 100 examples each, selected with ChatGPT assistance. Static covers static objects or camera motion, Articulated covers human or animal motion, and Interaction covers human–object or multi-object interactions.

R2. Results by scene category.

MethodSceneSubject consistency (SC)Background consistency (BC)Motion smoothness (MS)VBench Avg.RSPSL3RCGSSCPTCOverall RC
WanStatic94.1796.2998.6496.3744.3776.8787.3851.2066.19
+GeoRouteStatic98.4398.6298.9698.6754.2485.4789.0450.9669.03
WanArticulated93.0595.2198.4695.5742.8073.8985.3450.7264.72
+GeoRouteArticulated97.9797.8598.6798.1653.2983.1788.9652.0469.07
WanInteraction91.0292.5896.4493.3538.7669.3581.1745.6259.90
+GeoRouteInteraction97.4597.2898.6097.7853.1383.0288.7651.8968.90

The interaction subset has the largest RC improvement. GeoVideoBench evaluates the principal foreground object's local rigidity, long-range 3D consistency, global shape stability, and silhouette topology. In multi-object scenes, PTC can capture some topology changes visible in the retained foreground silhouette, including changes in holes, skeleton branches, and thin structures. The correctness of interactions themselves, including their semantics and physical dynamics, is outside the benchmark's scope.

AuthorsOpenReview ↗

GtiM — Response 2/4

Q2. Critical Weakness 2

The Spectral Router contributes negatively on the topology dimension. In the ablation, the full method scores lower on topology than three simpler variants, namely fixed routing, adaptive timestep, and window injection, and on Wan the improvement on this dimension is also the weakest among the four geometric dimensions in the main results; the paper neither analyzes this counter-intuitive result nor mentions it in the limitations. One possible explanation is that spectral injection is low-pass filtering and can hardly convey the mid- and high-frequency information needed to repair topological errors, but topological corruption is not necessarily determined by high-frequency information alone, and the paper provides no correspondence between frequency bands and corruption types to verify this, so the explanation remains a hypothesis. In addition, the key hyperparameters governing routing and injection are given directly, and the ablation includes no sensitivity analysis; a supplementary analysis is suggested.

Topology and injected frequency bands.

Thank you for pointing out the topology result. We investigated it by training two additional variants that retain the original adaptive low-frequency mask and add middle- or high-frequency information, respectively.

R3. Injection frequency bands.

BandsRSPSL3RCGSSCPTCRC
Low53.2083.3188.7651.8968.95
Low + Mid52.3182.8786.9450.1367.43
Low + High51.2880.0482.1952.3766.63

R3 shows an empirical trade-off: adding high-frequency information improves PTC but reduces the other geometric scores and Overall RC. We therefore retain low-frequency injection for its better overall balance.

Learned parameters and sensitivity.

The cutoff ρ\rho, boundary sharpness β\beta, and injection strength γ\gamma are predicted by the learned Spectral Router for each sample and selected timestep. Adapting both the mask and strength gives the best results in R4; changing the predicted strength by ±25% modestly lowers the aggregate scores but preserves the gains over Wan in R5.

The fixed settings in R4 are also the initial values for the corresponding adaptive outputs. The router predicts them from compatibility and timestep, with ranges ρ∈(0.08,0.24)\rho\in(0.08,0.24), β>4\beta>4, and γ∈(0,0.10)\gamma\in(0,0.10) (Appendices B.3 and D).

R4. Mask and strength adaptation.

Mask (ρ,β\rho,\beta)Strength (γ\gamma)RSPSL3RCGSSCPTC
AdaptiveAdaptive53.2083.3188.7651.89
AdaptiveFixed: γ=0.05\gamma=0.0551.2479.7186.9350.66
Fixed: ρ=0.18,β=16\rho=0.18,\beta=16Adaptive50.6980.1588.2950.98
Fixed: ρ=0.18,β=16\rho=0.18,\beta=16Fixed: γ=0.05\gamma=0.0544.7676.6388.5251.02

Joint adaptation achieves the highest score on all four geometric dimensions in this comparison. We also trained a single set of spectral parameters shared across all samples to separate the effect of sample adaptation from trainability. This shared control gives VB/RC of 97.73/68.63, compared with 98.15/68.95 for the sample-adaptive router; the shared control has higher PTC, consistent with the trade-off in R3.

R5. Injection-strength sensitivity. The predicted strength is scaled as γused=sγpred\gamma_{used}=s\gamma_{pred} with the checkpoint, prompts, seeds, and priors fixed.

ScaleSCBCMSVBRSPSL3RCGSSCPTCRC
0.7597.6497.2398.5297.8053.1782.9687.5451.6868.46
1.0097.9297.8998.6398.1553.2083.3188.7651.8968.95
1.2597.2297.1098.1797.5053.0183.1287.9551.6968.59

The default prediction gives the highest VB and RC. At both perturbed strengths, these aggregates remain above Wan's 95.61/65.36 in R9.

AuthorsOpenReview ↗

GtiM — Response 3/4

Q3. Critical Weakness 3

Results are reported as single point estimates, and the paper reports no variance, multiple random seeds, or significance tests, so the narrow margins over the strongest competitor can hardly support a stable ranking. Several results lack mechanistic explanation: the subject-consistency gain on one backbone far exceeds that on the other two, and single-time injection nevertheless increases the denoising time, for which the paper offers an attribution but no explanation; the selection criterion of the qualitative samples is also unstated. Furthermore, the human-alignment argument juxtaposes the proposed metrics with numbers reported by other works on different tasks and provides no inter-rater agreement reference, making the absolute level of the scores hard to calibrate.

Multiple seeds and statistical significance.

GeoRoute obtains higher VB and RC than VideoGPA in all five additional matched generation runs, and its improvements in VB and Motion Smoothness remain significant after Holm correction. The original single-seed comparison did not quantify variability across runs. On the shared CogVideoX backbone, we calculated aggregates separately for each seed and report sample standard deviations, paired 95% confidence intervals, and two-sided paired t-tests with Holm correction across all nine scores.

R6. Five-seed comparison. Δ is GeoRoute minus VideoGPA.

StatisticSCBCMSVBRSPSL3RCGSSCPTCRC
VideoGPA,mean±SD95.412±0.45595.594±0.40997.454±0.30096.153±0.15642.358±0.36675.224±0.64084.494±0.74051.876±1.25465.215±0.624
GeoRoute,mean±SD96.322±0.28995.790±0.12098.394±0.17696.835±0.14143.012±0.77176.172±0.10285.512±0.46452.194±0.66065.884±0.284
95% CI for Δ[0.057,1.763][-0.209,0.601][0.487,1.393][0.347,1.017][-0.473,1.781][0.126,1.770][0.049,1.987][-0.488,1.124][0.047,1.291]
Unadjusted p0.04150.25010.00450.00480.18230.03290.04330.33480.0405
Holm-adjusted p0.24280.54690.04040.04040.54690.23030.24280.54690.2428

The mean improvements are 0.682 in VB and 0.669 in RC. For RC, the unadjusted confidence interval is positive, while the Holm-adjusted p value is 0.2428. The five-seed result therefore supports a consistent observed RC gain, with stronger statistical evidence for VB. The separate prior-seed experiment in R1 shows that the result is also stable across the five tested prompt-paired 3D priors.

Backbone differences.

Wan's lower starting Subject Consistency provides more room for improvement and is a plausible explanation for its larger gain. In original Table 1, Wan starts at 93.18 and gains 4.74, compared with 94.89 and +0.17 for HunyuanVideo and 94.72 and +1.37 for CogVideoX.

Inference time.

We apologize for the confusion caused by our original timing breakdown. The denoising category included 3D-to-2D projection, tensor concatenation, and sparse/dense point-cloud conversion, which obscured the cost of denoising itself. R7 now reports these operations separately under routing/prior processing and includes online TRELLIS generation for every prompt. A single injection still requires prior generation, compatibility evaluation, and feature processing, which account for most of the additional time.

R7. Inference latency in seconds.

MethodTRELLISRouting/priorDenoisingOthersTotalOverhead
Wan2.1--203.916.08209.99-
Wan2.1 + GeoRoute7.0832.74204.175.89249.8819.00%
CogVideoX--211.3220.78232.10-
CogVideoX + GeoRoute9.0443.33212.9620.69286.0223.23%
HunyuanVideo--1174.8138.341213.15-
HunyuanVideo + GeoRoute17.1995.181175.2238.581326.179.32%

Denoising itself increases by only 0.26–1.64 seconds across the three backbones. Including prior generation and all processing, the end-to-end overhead is 9.32%–23.23%.

Qualitative sample selection.

The paper's qualitative method comparisons were randomly sampled from outputs generated with VBench and OpenVid prompts. Original Fig. 19 separately shows the bottom 3% by GeoVideoBench RC to illustrate failure cases. Website examples were manually selected to make geometric artifacts and method differences easier to inspect.

Comments on the paper

1 comment
AuthorsOpenReview ↗

For All Reviewers

Thank you for your careful reviews and for recognizing GeoRoute's empirical motivation, controlled ablations, benchmark validation, and results across video backbones. We have added the following evidence in response to your questions:

  1. Multi-seed comparison. The reported VBench average and GeoVideoBench RC improve over VideoGPA in all five paired runs.

  2. Independent evaluation. External geometric metrics and automated VideoReward preferences favor GeoRoute over VideoGPA. GeoRoute also improves all five reported PDI-Bench metrics over Wan.

  3. Dynamic scenes. Gains extend to articulated subjects, human–object and multi-object interactions, and all baseline motion groups. Dynamic Degree does not decrease on any of the three backbones.

  4. Adaptive routing. Learned shared-parameter controls, timestep and frequency-band tests, strength perturbations, and direct comparisons with related methods provide further evidence for the adaptive single-injection design.

The individual responses provide the detailed results and clarify inference inputs, benchmark scope, inter-rater agreement, and computational cost.

The corresponding revisions in the main text and appendices are highlighted in blue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.