acceptodds
Under review as a conference paper at ICLR 2027

AO-FLOW: LEARNING TO USE AND OPTIMIZE GENERATED VIEWS FOR 3D ASSETS

Abstract

Orbit videos provide candidate observations for single-image 3D generation, yet different regions need different views and adjacent frames can repeat the same error. AO-Flow first adapts a 3D receiver to use the full 49-frame video: sparse- structure sampling retains region-specific source identities and attention masses, and Shape rereads those frames with its own queries and a soft spatial prior. Se- quential supervision trains the source producer before freezing it for Shape adapta- tion. AO-Flow then fixes the receiver and uses final-asset feedback to update only video attention LoRA. The regional interface reaches H-F 0.639 versus 0.613 for a global four-source reader at similar receiver cost; full 49-frame reading attains slightly higher H-F at greater cost. Asset feedback raises H-F to 0.669 while trad- ing some input-view agreement. Inference uses one image, one video and one unselected asset

Primary area:
generative models
Keywords:
IMAGE-TO-3D, POST-TRAINING, RL

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_M6XTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents AO-Flow, a framework for improving single-image 3D asset generation by better utilizing and optimizing video-generated multi-view observations. The proposed method consists of two main components. First, it introduces a cross-stage regional conditioning mechanism for TRELLIS.2. The Sparse Structure (SS) stage processes 49 generated video frames and records region-specific source identities and attention weights, which are subsequently reused by the Shape stage with its own content attention and spatial priors. The two stages are adapted sequentially to ensure consistent information transfer. Second, the authors introduce an asset-oriented video post-training strategy based on Flow-GRPO. By freezing the adapted 3D receiver and optimizing the video generator using final 3D asset rewards, the framework aims to improve geometric completeness and hidden-surface reconstruction. Experiments on Objaverse-derived objects and GSO demonstrate improvements in geometric reconstruction metrics, particularly hidden-region F-score, compared with several existing 3D generation approaches.

Critical strengths
  • The paper is well-written and easy to follow.
  • The paper addresses a practically relevant problem by preserving region-specific multi-view evidence for fine-grained 3D reconstruction.
  • The framework coherently combines regional source selection, sequential receiver adaptation, and asset-level video optimization guided by final reconstruction quality.
  • Extensive comparisons, ablations, and efficiency analyses provide useful insights into source selection, spatial priors, training objectives, geometry, and input-view preservation.
Critical weaknesses
  • The methodological novelty and specific contribution of cross-stage regional memory remain unclear, as gains over alternative routing strategies are small and may partly result from retained attention mass or spatial priors.
  • Statistical reliability is insufficient because several conclusions rely on small improvements without multi-seed results, confidence intervals, or significance tests.
  • Asset Flow-GRPO provides only marginal gains over Asset-DPO despite higher computational cost and slightly worse input-view preservation, weakening its practical justification.
  • Generalization to real photographs and imperfect generated views remains underexplored, particularly under viewpoint drift and geometric inconsistencies.
  • The complete pipeline’s computational overhead is unclear because reported efficiency excludes video generation and other shared stages; an end-to-end comparison is needed.
  • Figures 2 and 3 are overly dense and difficult to read. Clearer vector-based diagrams, larger text, and simplified layouts would improve technical communication.
Questions for authors
  • Could the authors better isolate the contribution of cross-stage source memory from attention mass, spatial bias, and source allocation through matched controls?
  • Are the improvements of Region-4 over Shape-selected-4 and Flow-GRPO over Asset-DPO statistically significant across multiple seeds and inference draws?
  • What practical advantage does Flow-GRPO offer over the cheaper Asset-DPO? Please provide compute-matched comparisons and analyze reconstruction–fidelity trade-offs.
  • How robust is regional source selection to camera errors, temporal inconsistencies, and hallucinated geometry? Would training with generated videos improve robustness?
  • Could the authors report complete image-to-3D runtime and peak memory, including video generation and encoding, and evaluate on real photographs?
  • Could Figures 2 and 3 be redesigned with larger text, simpler layouts, and clearer data flows and optimization scopes?
2 comments
AuthorsOpenReview ↗

Response to Reviewer M6XT: results for the six requested controls

Thank you for the concrete list of controls. We ran all six. Unless marked single-seed, H-F averages 3 training seeds × 4 inference draws, with paired bootstrap 95% CIs over 454 objects, so values can differ slightly from single-run tables.

Q1 (W1). Contribution of cross-stage memory

Each row removes one factor from Region-4:

ControlFactor removedH-FΔ
Region-4–0.640–
IDs shuffled across regionsRegion–frame match0.618−0.022
Global-4 IDs, Region massRegion-specific frames0.620−0.020
No spatial biasSpatial prior0.625−0.015
Uniform weightsRelative allocation0.631−0.009
Region-4 IDs, Global massRetained mass0.632−0.008
Shape re-selects sourcesCross-stage reuse0.634−0.006

Region-specific frame identity is the largest factor, ahead of spatial bias, allocation and mass. Removing the bias from both Global-4 and Region-4 leaves a 0.022 gap, so it does not explain the regional gain. These per-region mass rows replace Table 12 (reply to L12K). Region-4 adds 0.8 s of receiver time over Global-4 (Table 2).

UMI3D, ROAR-3D and Stream3D choose views within one stage or with a separate probe. AO-Flow reuses the choice SS makes on its 16³ grid as the prior for Shape's 32³ cells, which the last row isolates. Against Shape-selected-4, which probes all 49 frames at every Shape step, reuse gains 0.006 H-F at 98.2 vs. 118.6 s and 32.0 vs. 38.4 GB.

Q2 (W2). Significance

Region-4 over Shape-selected-4 excludes zero, as do the other interface, reward and system contrasts. GRPO over DPO does not, and Q3 shows where they differ in practice.

ContrastΔ H-F95% CI
Region-4 vs. Global-4+0.024[0.019, 0.029]
Region-4 vs. Shape-selected-4+0.006[0.002, 0.010]
Asset GRPO vs. geometry-video RL+0.007[0.003, 0.011]
Asset GRPO vs. Asset-DPO+0.003[−0.001, 0.007]
AO-Flow vs. Pixal3D+0.020[0.013, 0.027]

The revised tables report these means and CIs.

Q3 (W3). Flow-GRPO vs. Asset-DPO

Rows below 1,000 updates are checkpoints of the same GRPO runs.

Policyλ_oGPU-hH-FInput IoU
Asset-DPO27900.6660.910
Asset GRPO, 600 upd.25180.6640.908
Asset GRPO, 900 upd.27780.6670.907
Asset GRPO, 1000 upd.28640.6690.906
Asset GRPO, 1000 upd.48640.6660.911

At equal compute the two are level (0.667 at 778 GPU-h vs. 0.666 at 790). They differ in when compute pays off. DPO spends 690 of its 790 GPU-h on its pool before the first update (Table 13), while GRPO improves throughout. GRPO reaches 0.664 at 518 GPU-h, before DPO would have started training.

The preservation weight λ_o sets the reconstruction–fidelity trade-off. From λ_o = 0 (Table 4) to 4, input IoU rises from 0.873 to 0.911 while H-F falls from 0.678 to 0.666, and at λ_o = 4 GRPO matches DPO on both.

Q4 (W4). Robustness and training on generated videos

Region-4 leads under each error type, and tuning Shape on generated videos helps most where videos are worst. Single-seed, same runs as our L12K and DHkb replies:

ConditionGlobal-4Region-4+ Shape tuned on generated
Clean0.6130.6390.647
Pose error > 15° (11% of videos)0.5850.6040.621
8 frames from another video0.5840.6250.637

Tuning adds 0.017 under large drift and 0.008 on clean videos, where the same budget on GT renders gives 0.642 (Table 10). For hallucinated geometry, a wrong part repeated in 4 adjacent frames (8% of the video) drops to 2% of Shape's sources, and Region-4 degrades no more than Full-49 (reply to DHkb).

Q5 (W4, W5). End-to-end cost and real photographs

Table 2's receiver times run from DINO encoding to mesh export (App. I), so we now add video generation (one H200, batch one). Real photos are 200 OmniObject3D objects, with the scan as reference.

MethodTime (s)Peak GBH-F ObjaverseH-F real
TRELLIS.238210.5240.462
Pixal3D52260.6490.566
AO-Flow, 8-step video167680.6630.576
AO-Flow270680.6690.581

On real photos AO-Flow leads Pixal3D, the strongest Table 1 baseline, by 0.015 H-F, and the 8-step sampler keeps the lead at 38% less runtime. The extra time over single-stage generators is mostly video generation, which takes 64% of runtime and sets the 68 GB peak (the receiver needs 32 GB). The table is now in Sec. 4, and 50 in-the-wild results are at 4:20 in the video.

Q6 (W6). Figures 2 and 3

Both are redrawn as vector graphics with text of at least 8 pt. Fig. 2 shows S1, S2a → S2b and S3 as separate panels and marks what each stage trains. Fig. 3 drops the repeated 49-frame strip and traces one path from SS capture to Shape rereading.

All six points are addressed in the revision. If these results resolve your concerns, we would welcome an updated rating.

Reviewer_M6XTOpenReview ↗

Reply to Authors

Thank you for the additional experiments. I found the comparison with Shape-selected-4 particularly helpful. Reusing the sources selected by SS gives a small gain in hidden-region F-score while reducing both runtime and memory. This addresses my main concern about the value of cross-stage reuse.

I also read the other reviewers’ comments and share the concern raised by Reviewer DHKb about the reliability of generated views as a geometric prior. The robustness tests and real-photo results are encouraging. However, the real-photo comparison evaluates the full system, so it remains difficult to tell how much of the gain comes from the generated views and how much comes from adapting the TRELLIS.2 receiver. I also found the compute-matched comparison informative. Flow-GRPO produces a useful checkpoint earlier, but its final advantage over Asset-DPO remains uncertain.

The rebuttal has strengthened my assessment of the regional conditioning method. My remaining concerns about the video prior and the added value of Flow-GRPO lead me to maintain my original rating of 2 (weak rejection).

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

This paper presents AO-Flow, a single-image 3D generation framework that uses a generated 49-frame orbit video as additional evidence. It introduces a regional cross-stage interface that transfers view identities and attention priors from the Sparse Structure stage to the Shape stage. The method further optimizes the video generator with final-asset rewards while keeping the 3D receiver fixed. Experiments show improved hidden-surface reconstruction and competitive overall geometry quality, with a trade-off in input-view fidelity and computational cost.

Critical strengths
  1. The regional view-selection mechanism is intuitive and allows different 3D regions to use different supporting views while avoiding the full cost of reading all 49 frames.

  2. Optimizing the video generator using final 3D asset quality provides better alignment with the downstream objective than relying only on video-level supervision.

  3. The paper includes extensive ablations on the regional interface, reward components, alternative post-training strategies, and computational costs.

Critical weaknesses
  1. The framework and result figures appear distorted when enlarged. They contain excessive white space and too many fine-grained details, while the tangled connection lines make the figures difficult to follow and negatively affect the reader’s first impression.

  2. The motivation is not sufficiently convincing. Existing methods can already reconstruct a 3D asset directly from an orbit video, so it is unclear why the generated video needs to pass through the TRELLIS architecture. If the generated orbit video is inconsistent across views, using it may be less reliable than applying TRELLIS directly to the single input image. Conversely, if the video is sufficiently consistent, why not reconstruct the 3D asset directly from it? Moreover, there is a training–inference gap: training uses ground-truth orbit trajectories, whereas inference relies on generated trajectories.

  3. The interpretation of SS attention mass as reliable regional evidence is insufficiently validated. High attention does not necessarily imply that a view is geometrically correct or informative. In particular, neighboring video frames may repeat the same hallucinated structure, in which case the regional memory could consistently route correlated errors into the Shape stage. The current ablations show that the proposed routing mechanism is useful on average, but they do not demonstrate that it can distinguish reliable observations from repeated video-generation artifacts.

  4. The spatial conditioning assumes that the generated video follows the nominal camera trajectory, but this assumption may not hold in practice. A video diffusion model may partially rotate or deform the object, change its scale, or otherwise deviate from the prescribed camera motion. In such cases, projecting a 3D query using the nominal camera can associate it with an incorrect image region.

  5. The asset-level post-training may optimize the video for a particular frozen receiver rather than improve its general geometric quality. Because the 3D receiver remains fixed throughout Stage 3, the video policy may learn receiver-specific visual cues that increase the downstream reward without producing more realistic or geometrically consistent views.

Questions for authors
  1. The submission would benefit from a supplementary video and additional qualitative visualizations that enable a more intuitive comparison of the generated views, cross-view consistency, and final 3D assets. Moreover, the current supplementary file appears to largely reproduce the main manuscript rather than provide substantive additional material.
2 comments
AuthorsOpenReview ↗

Response to Reviewer DHkb: failure-case tests on generated videos

Thank you for the detailed review. Most of your concerns ask whether an assumption holds on generated videos, so we tested each failure case directly.

W2. Motivation and the train–test gap

In practice most videos fall between your two cases. In the median video, 87% of frames agree with a VGGT reconstruction within 5% depth error. We split test objects into thirds by this error, with the same videos for every method:

MethodAllMost consistentLeast consistent
TRELLIS.2 image-only, tuned on our data0.5710.5830.556
2DGS, 49 frames, VGGT poses0.5480.6270.452
SV3D's NeRF-to-mesh pipeline0.5610.6410.468
AO-Flow S2 (Region-4)0.6390.6680.603

Direct reconstruction beats image-only on the most consistent third and falls well below it on the least consistent, as you expected, while AO-Flow leads in every third. On consistent videos the 3D prior still fills surfaces the orbit never sees, such as the underside. On inconsistent ones regions shift weight to null, rising from 0.31 to 0.47 across the thirds. AO-Flow matches or beats image-only on 91% of objects.

S2 trains on GT renders, and switching to generated videos at test costs 0.062 H-F (0.701 vs. 0.639). Tuning Shape on generated videos beats the same budget on GT renders (0.647 vs. 0.642, Table 10). Asset-RL videos reach 0.669, and both together reach 0.675, closing 58% of the gap.

W3. SS attention mass and repeated artifacts

We agree that attention is not correctness, and Sec. 3.2 now calls SS mass a routing prior. Where GT renders can check it, SS mass separates frames with correct and wrong local depth at AUROC 0.69 (angular distance: 0.61). Extending the repetition panel of App. G.5, we inserted or deleted a part in k adjacent frames of 20 validation objects. Local F-score over both edit types:

Receiver \ k adjacent frames (% of video)01 (2%)4 (8%)8 (16%)
Global-40.580.550.500.45
Full-490.650.640.620.53
Region-40.640.630.610.53
Corrupted share of sources, SS → Shape–0.4% → 0.2%3% → 2%12% → 7%

Region-4 degrades no more than reading all 49 frames, so regional memory does not amplify repeated errors. Short runs are down-weighted by SS and again by Shape. At k = 8 the wrong part is seen consistently from viewpoints spanning about 55°, leaving the receiver no conflicting view to check it against. This is the case Stage 3 targets. Asset RL cuts videos with annotated repeated artifacts from 23% to 14%, and Limitations now discuss the rest.

W4. Nominal cameras

Nominal cameras enter the reader keys and Shape's projection bias, which is bounded and never crops patches. Without the bias Region-4 still beats Global-4 (0.624 vs. 0.613, Table 2), so most of its gain does not rely on the projection. VGGT finds 11% of generated videos drifting more than 15°.

Using VGGT cameras instead of nominal ones adds 0.012 H-F on those videos and about 0.002 overall, so the nominal assumption costs little. Region-4 still leads Global-4 there (0.604 vs. 0.585), and VGGT cameras are now an option (+4 s). ±10% scale jitter lowers H-F by 0.003. A smooth random rotation and warp of the object in 8 frames costs Region-4 0.011 and Global-4 0.024.

W5. Receiver-specific optimization

Most of the asset-RL gain carries over to receivers never used in training. We fed Region-4 asset-RL videos, without retraining, to other receivers:

ReceiverInitial videoAsset-RL videoGain (share kept)
Region-4 (training receiver)0.6390.669+0.030
Global-4 (same SS stage)0.6130.637+0.024 (80%)
TRELLIS.2 multi-image0.5820.605+0.023 (77%)
2DGS (no learned prior)0.5480.569+0.021 (70%)

Receivers with no parameters trained by us keep 70–77% of the gain, and the rest is specific to the training receiver. On the videos, VGGT reprojection error falls by 19% (0.081 to 0.066), while frame FID vs. GT renders (31.4 vs. 30.9) and DINO similarity to the input (0.842 vs. 0.846) barely change. VGGT is not in the asset reward. The gain also survives JPEG-75 (94% kept) and 1-pixel blur (91%), so it is not carried by fine pixel patterns.

W1 and Q. Figures and supplementary (later)

Figs. 2, 3 and 7 are redrawn as vector graphics with larger text and no crossing connectors, and Figs. 1, 4, 5, 8 and 9 use higher-resolution renders with less white space. The supplementary PDF now holds only new material. Its video shows orbits, VGGT point clouds for cross-view consistency and turntables for 40 test objects, with a repeated artifact , pose drift, initial vs. asset-RL orbits and per-region retained frames.

Two results now appear in the paper: a partly inconsistent orbit stays useful when each region can fall back on the image prior, and asset feedback makes videos more consistent for other receivers. We hope you will weigh them in your final rating.

There is no video in supplementary materials. Did I miss it? I only find a pdf paper no differences with the original paper.

Reviewer_L12KOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AO-Flow, a framework for improving single-image 3D asset generation by better utilizing generated multi-view observations and optimizing the video generator with downstream 3D feedback. Specifically, the authors first use a video diffusion model to generate a 49-frame orbit video from a single input image. To better utilize these generated views, they propose a regional conditioning mechanism for TRELLIS.2, where the Sparse Structure stage selects informative views for different spatial regions and transfers the selected view information to the Shape generation stage. The Shape stage then uses these selected views with its own attention and spatial guidance to generate more detailed geometry. Furthermore, the authors introduce an asset-oriented reinforcement learning stage, which uses the quality of final reconstructed 3D assets as reward to optimize the video generator through Flow-GRPO. Experiments demonstrate improvements in 3D reconstruction quality, especially for unseen regions.

Critical strengths
  1. The paper addresses an interesting and meaningful problem on how to better utilize generated multi-view observations for improving 3D reconstruction quality.

  2. The proposed cross-stage regional conditioning is an interesting design, which allows different 3D regions to select their own informative views. It achieves comparable reconstruction quality to using all frames while reducing the computation cost.

  3. The experiments are relatively comprehensive, including comparisons with several existing methods and ablation studies on different components of the proposed method, which help to validate the effectiveness of their design.

Critical weaknesses
  1. My main concern lies on the effectiveness and computational cost of the proposed RL training. Since each generated video needs to go through the entire 3D generation pipeline to calculate the reward, I believe this would introduce significant training overhead. As shown in Table 3, Asset Flow-GRPO requires 864 GPU-hours, compared with 552 GPU-hours for Geometry-video RL, while only improving H-F from 0.663 to 0.669. I am not fully convinced whether the additional computational cost of online RL is justified by such relatively small improvement.

  2. In Table 2, Region-4 improves H-F from 0.613 to 0.639 compared with Global-4. However, the improvement reduces to only 0.006 when matching the retained attention-prior mass (Appendix Table 12). This makes me question how much of the improvement actually comes from the proposed regional source selection, rather than the difference in attention allocation.

Questions for authors
  1. I understand that the authors have discussed the motivation of using video diffusion model instead of multi-view image generation. However, I wonder whether this advantage can be further validated through a controlled experiment. For example, using the same TRELLIS.2 receiver while changing the view generator between video diffusion and multi-view image diffusion. Multi-view image diffusion models can generate fewer but more sparsely distributed viewpoints, which may provide sufficient geometric information with lower computational cost. I think such comparison would be helpful to better demonstrate the benefit of using video generation.

  2. How sensitive is the proposed regional view selection to the inconsistency and camera pose errors in generated videos? I wonder how well the proposed method can handle inaccurate geometry or camera poses in the generated views.

1 comment
AuthorsOpenReview ↗

Response to Reviewer L12K: compute-matched RL comparison and attribution of the regional gain

Thank you for the careful review. In short, most of the Region-4 gain comes from which frames each region keeps, asset RL is level with geometry-video RL at equal compute and pulls ahead with more updates, and with the same receiver, video beats multi-view diffusion mainly through frame density. H-F values are 3-seed means unless marked single-seed, and can differ slightly from the single-run paper tables.

W1. Cost of asset RL

Geometry-video RL is also online Flow-GRPO with the same update rule (Sec. 4.3). Most of the 312 GPU-h difference is the cost of turning each candidate into an asset and scoring it (candidates 380 vs. 280, scoring 320 vs. 170 GPU-h, Table 13). For equal compute we evaluated the saved checkpoints tracked in Fig. 10 on the test set.

PolicyGPU-hH-FStruct. pass
Best-of-4 SFT4260.65572%
Asset GRPO, 500 upd.4320.66178%
Asset GRPO, 600 upd.5180.66479%
Geometry-video RL5520.66277%
Asset GRPO, 900 upd.7780.66781%
Asset-DPO7900.66680%
Asset GRPO, 1000 upd.8640.66982%

Asset GRPO is level with geometry-video RL by 518 GPU-h, and the rest of the budget buys +0.007 H-F (95% CI [0.003, 0.011]). H-F averages over the whole hidden surface, so a missing handle or detached leg barely moves it. On the blinded structure checklist, which no reward uses, failures fall from 23% to 18% of objects, about a fifth fewer. Examples are at 3:20 in the supplementary video.

Training offline does not avoid this cost. DPO spends 690 of its 790 GPU-h on its candidate pool before the first update (Table 13) and reaches 0.666, against 0.667 for GRPO at the same compute. Both beat Best-of-4 SFT, which only imitates the winner, by 0.011 to 0.014. None of them change inference cost.

W2. Frame selection vs. prior mass

Most of the gap comes from which frames each region keeps. A global set shares four frames across all regions, so it also keeps less of each region's own mass. To separate the two, we crossed frame IDs with mass level, matched region by region, scaling Global priors up for this diagnostic only.

Frame IDs \ Prior massGlobal levelRegion levelMass effect
Global-4 IDs0.6160.620+0.004
Region-4 IDs0.6320.640+0.008
Frame-ID effect+0.016+0.020

At equal mass in every region, switching to Region-4 frame IDs adds 0.016 to 0.020 of the 0.024 gap, while more mass with the same IDs adds 0.004 to 0.008. Giving each region another region's frame IDs while keeping its own masses yields 0.618, close to the Global-IDs row. Table 12 equalized the grid-wide total (0.72) in a single run, so per-region mass still differed between the arms. The per-region version replaces it in Sec. 4.2.

Q1. Video vs. multi-view diffusion

All sources use the same frozen S2 receiver (single seed), adapted on GT renders, so it saw neither generator's output. MV-Adapter generates 6 views at our 15° orbit elevation, and the middle row reads the video frames at the same 6 azimuths but pays for all 49. At MV-Adapter's native 0° elevation, H-F is 0.608.

View sourceGenerated / readH-FInput IoUEnd-to-end (s)
MV-Adapter6 / 60.6110.91285
Video49 / 60.6180.910222
Video, Region-449 / 490.6390.911270

With the same six cameras, video frames give 0.007 higher H-F, and reading all 49 adds another 0.021, mostly on thin parts and openings, so the advantage comes mainly from frame density. MV-Adapter is about 3x faster. An 8-step video sampler brings our pipeline to 167 s (reply to M6XT), and the paper now reports end-to-end time. Side-by-side assets are at 2:50 in the video.

Q2. Pose error and inconsistent frames

On generated videos, binned by VGGT pose error (single seed):

Pose errorObjectsGlobal-4Region-4+ VGGT cameras
< 5°58%0.6210.6490.649
5–15°31%0.6070.6320.635
> 15°11%0.5850.6040.616

Region-4 keeps its lead in every bin. On the 11% that drift most, VGGT cameras add 0.012 for 4 s, and we include them as an option. This fits Fig. 6b, where the projection bias helps up to about 10° of azimuth offset and is about 0.02 below the no-bias reader at 20°. Swapping 8 of the 49 frames for frames from another video of the same input lowers Region-4 by 0.014 and Global-4 by 0.029, as they are 16% of the video but 5% of retained sources. A wrong part repeated in 4 adjacent frames (8% of the video) falls to 2% of Shape's sources, and Region-4 degrades no more than Full-49 (reply to DHkb). Pose-drift cases are at 1:25 in the video.

We also removed internal labels such as π₀¹³ and moved the attribution tables into the main text. We hope the matched-compute comparison and the frame-identity control resolve your two main concerns, and would appreciate it if you would reconsider your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.