Literature
Display
← All papers

LoGo: Local-Global Rewards for Consistent Long-Horizon Video Generation

Ziqi Ma · Shreya Sharma · Mohamed El Banani · Katja Schwarz · Chongjie Ye · Chao-Yuan Wu · Li Fei-Fei · Ben Mildenhall · Georgia Gkioxari · Justin Johnson · Gowthami Somepalli

Preprint posted · arXiv (preprint) ↗

Computer vision11 min text · CondensedOriginal paper ↗
Contents

Overview

This edition follows the complete 17-page v1, retaining all ten numbered figures, eleven tables, four displayed equation blocks and 45 references. Figures 6 and 7 share an HTML container in the source; they are separated here so both original captions and all panels remain visible.

Complete original PDF block S0.F1
Figure 1: Camera-controlled video models often break 3D consistency during generation (top row). Here, a video generated by an existing video model shows both local inconsistencies (the numbers on the wall are distorted) and global inconsistencies (the wall’s orientation shifts). We introduce LoGo, a post-training approach that combines global and spatially localized rewards to reliably improve consistency in generation (bottom row).

At a glance

LoGo post-trains camera-controlled video generators with a combination of voxel-localized and whole-video geometry rewards. Local feedback gives individual scene regions their own training weight; global feedback and interleaved aesthetic/camera rewards counter the quality losses observed with local geometry alone.

The experiments span three 14B generators, two benchmarks and three optimization methods. The main evidence is improved reconstruction-based consistency on static scenes. Very long videos, dynamic worlds and uniform gains across every camera or quality metric are not established.

1 Introduction

A stable generated world needs both a persistent scene layout and objects that retain their geometry and appearance when revisited. Longer paths expose local hallucinations, disappearing objects and structural changes that a whole-video average can dilute. LoGo tests whether assigning feedback to 3D regions improves this credit assignment.

The scalar-reward failure in Figure 2 is an empirical limitation of the tested global average, not a theorem that every possible scalar objective cannot penalize a local error. The contribution is a practical local/global weighting recipe and a benchmark with longer, more varied camera paths.

Complete original PDF block S1.F2
Figure 2: A scalar reward cannot capture localized failures: the door inconsistency is “washed out” by the overall good structure of the rollout. LoGo catches the localized inconsistency (right).

2 Related Work

Earlier geometry-based post-training uses epipolar consistency, depth reprojection or Gaussian reconstruction. Architectural geometry, memory and joint geometry/video modeling offer complementary approaches. Many prior reward experiments concern short, 80-frame generators.

Spatial reward or loss masking also appears in image editing and robotic video generation. LoGo focuses on matching training credit to persistent 3D regions over longer trajectories. Camera-following or action-reversal objectives address related properties but do not by themselves establish scene consistency.

3 Method

The method has three components: aggregate reconstruction errors in voxels, blend normalized local and global feedback, and alternate geometry updates with aesthetic and camera updates. The recipe changes post-training supervision rather than adding a new persistent world representation to the generator.

3.1 Reward localization in 3D

Keyframes pass through VGGT to estimate depth and camera geometry. Pixels are unprojected into a common scene cloud and grouped into voxels. Equation 1 averages a weighted squared RGB reprojection error and an absolute depth reprojection error over the pixels assigned to each voxel. The depth term is absolute, not squared.

The default voxel edge length is 0.1 times the scene-depth 90th percentile. A 240-frame video yields about 3,000 voxels per scene and 200 points per voxel at the reported medians. Optional cross-voxel normalization is mentioned for small rollout groups. The predicted geometry and correspondence quality limit the reliability of this local feedback.

The heatmaps highlight changing plant appearance, a newly appearing bookshelf and a local artifact. These examples show where the estimated reward places credit; they do not independently prove that the reconstruction has recovered the true scene.

Complete original PDF block S3.F3
Figure 3: Motivation of LoGo. (a) A global reward fails to improve 3D consistency; local reward improves consistency, measured in PSNR, but degrades video quality, measured in HPSv3. This motivates combining the two in LoGo. (b) LoGo (green) extends the Pareto frontier in the consistency-quality tradeoff. (c) Reward curves and quality evaluation during training: local reward degrades video quality over time; LoGo improves consistency without compromising quality.
Equation 1
ei=1|𝒮i|​∑(x,j)∈𝒮iα​‖ProjRGB​[𝒫,cj,x]−IjRGB​(x)‖22+(1−α)​|ProjD​[𝒫,cj,x]−IjD​(x)|\displaystyle e_{i}=\frac{1}{|\mathcal{S}_{i}|}\sum_{(x,j)\in\mathcal{S}_{i}}\alpha\left\lVert\text{Proj}_{\text{RGB}}[\mathcal{P},c_{j},x]-I_{j}^{\text{RGB}}{(x)}\right\rVert_{2}^{2}+(1-\alpha)\left|\text{Proj}_{\text{D}}[\mathcal{P},c_{j},x]-I_{j}^{\text{D}}(x)\right| (1)
Complete original PDF block S3.F4
Figure 4: a): illustration of voxel reward aggregation. b): visualization of local reward projected to video frames.

3.2 LoGo: Blending local and global reward

Local-only training improves measured consistency but can produce blur and muted colors. The authors interpret this as reward hacking: smoother edges and less color variation can reduce reconstruction error without preserving visual quality. Whole-video reward empirically suffers less of this degradation.

Equation 2 blends clipped, normalized improvements over local and global mean errors. Each contribution has coefficient 1/4 around a 1/2 offset, keeping the resulting optimality probability in [0,1]. This mixes normalized feedback rather than simply adding raw RGB/depth errors. The observed Pareto curves motivate the blend; the paper does not prove that it prevents all reward hacking.

The normalization symbols Z and group-comparison conventions are not fully specified in the text. Their exact implementation should not be inferred to be a particular standard-deviation formula.

Equation 2
ri,LoGo=12+14​clip​(e¯i,local−ei,localZei,local,−1, 1)+14​clip​(e¯global−eglobalZeglobal,−1, 1)\displaystyle r_{i,\text{LoGo}}=\tfrac{1}{2}+\tfrac{1}{4}\,\mathrm{clip}\!\left(\frac{\bar{e}_{i,\text{local}}-e_{i,\text{local}}}{Z_{e_{i,\text{local}}}},\,-1,\,1\right)+\tfrac{1}{4}\,\mathrm{clip}\!\left(\frac{\bar{e}_{\text{global}}-e_{\text{global}}}{Z_{e_{\text{global}}}},\,-1,\,1\right) (2)

3.3 Reward interleaving

Training cycles through k geometry steps, m HPSv3 aesthetic steps and n camera-reward steps. The camera objective combines rotation and translation errors with equal weights. Exact k:m:n values and the conversion/normalization between those errors are not supplied in this version.

The main optimizer is DiffusionNFT, with Flow-GRPO and DRaFT ablations. Computing pixel-to-voxel correspondence adds about 9 seconds to a 1,920-second training step on 64 H100 GPUs: roughly 0.5% of step time and 2.1% of reward time. This is incremental overhead, not the full cost of generating videos, reconstruction and backpropagation.

4 Experiments

Evaluation combines the new TrajectoryBench camera paths with a camera-conditioned adaptation of DL3DV. Comparisons include original generators, two adapted prior post-training methods and a global-only version of the authors’ reward.

4.1 Benchmarks

TrajectoryBench has 2,000 initial-image/trajectory tasks across easy, medium and hard levels, covering indoor, outdoor and doorway-transition scenes. Easy uses one motion over 80 frames from WorldScore; medium combines three motions in a continuous 240-frame sequence; hard uses seven or eight motions over up to 240 frames. Paths test expansion, revisits and entry into new spaces.

The DL3DV evaluation uses the prior 100-scene holdout but adds the original camera trajectories as conditioning. Results therefore concern this camera-controlled protocol and should not be equated directly with earlier image/text-only evaluation.

Complete original PDF block S4.F5
Figure 5: Camera trajectories of TrajectoryBench vs. WorldScore.

4.2 Training and evaluation settings

All three base models have 14B parameters. Lingbot-world-v2 generates 12-frame autoregressive chunks with cached attention; Lyra-2 generates 80-frame chunks with a sparse 3D cache; UniWorld is bidirectional and trained for 80-frame clips. Longer UniWorld trajectories are subsampled into 80 generated frames, so its path difficulty does not imply an equally long output horizon.

Post-training uses 1,700 DL3DV scenes and 16–36 hours on 64 H100s. VideoGPA’s scalar reprojection DPO and World-R1’s Gaussian/VLM Flow-GRPO are adapted from a different, shorter generator setting; the paper’s strongest credit-assignment comparison is its own global-only variant.

Geometry metrics include RGBD reprojection PSNR using VGGT or Depth Anything 3, depth consistency MVCS, Gaussian reconstruction PSNR and epipolar error. Camera RPE compares estimated motion to the commanded motion. VideoReward VQ is a separate evaluation score from the HPSv3 training reward. These are estimated proxy measures, not direct ground-truth geometry or human success judgments.

4.3 Results

On TrajectoryBench, Lingbot2’s PSNR-D rises 16.0→18.2 dB and epipolar error falls 1.507→1.111, about 26%. On DL3DV the corresponding changes are 13.1→15.8 dB and 2.161→1.365, about 37%. The three models improve the overall geometry columns, and native figures retain the examples of repaired objects, floaters and scene changes.

Camera and visual-quality improvements are not universal. UniWorld’s TrajectoryBench rotation error changes 13.8→14.0 degrees; Lyra2’s translation error changes .100→.101. On DL3DV, Lyra2’s LoGo VQ score −.513 is below VideoGPA’s −.508, and some baseline camera scores are stronger. These exceptions qualify broad preservation or best-result language.

The hard split often benefits more: Lingbot2 PSNR-D gains 1.0 dB on easy and 2.5 dB on hard. Appendix Table 6 provides the full breakdown and paired standard errors. Its claim of better performance in every metric is broader than the cells: Lyra2 easy MVCS is .922 versus the base .924, for example. Qualitative examples support the proposed mechanism without showing that every such failure is eliminated.

Complete original PDF block S4.T1
Table 1: Results on TrajectoryBench. PSNR-V and PSNR-D measure RGBD reprojection PSNR (using VGGT and DA3), PSNR-GS is based on Gaussian reconstruction. LoGo achieves strong gains in all consistency metrics across three base models. It increases PSNR-D by 2.2dB from Lingbot2 base model, and reduces epipolar error by 26%. Meanwhile, LoGo preserves, and often improves, camera following and video quality. Standard errors are in Table 6 in the Appendix.
Original selectable table
Model 3D consistency Camera Video quality
PSNR-V ↑\uparrow PSNR-D ↑\uparrow PSNR-GS ↑\uparrow MVCS ↑\uparrow Epipolar ↓\downarrow RPE∘ ↓\downarrow RPE trans ↓\downarrow VR-VQ ↑\uparrow
Lingbot2 Base 16.9 16.0 16.9 0.842 1.507 13.2 0.108 0.21
VideoGPA 17.1 16.1 17.0 0.847 1.486 13.1 0.107 0.22
World-R1 17.0 16.0 16.9 0.839 1.515 13.1 0.108 0.21
Global-only (ours) 18.0 17.1 17.8 0.872 1.306 11.2 0.102 0.31
LoGo (ours) 19.0 18.2 18.4 0.894 1.111 8.9 0.103 0.37
UniWorld Base 17.6 16.8 17.5 0.881 1.260 13.8 0.132 0.39
VideoGPA 17.7 16.9 17.5 0.882 1.275 13.8 0.132 0.39
World-R1 17.9 17.1 18.2 0.869 1.401 13.1 0.129 0.22
Global-only (ours) 18.3 17.5 18.0 0.892 1.195 13.3 0.126 0.37
LoGo (ours) 18.8 18.0 18.6 0.901 1.108 14.0 0.130 0.41
Lyra2 Base 18.4 17.7 19.0 0.897 1.550 7.9 0.100 -0.23
VideoGPA 18.0 17.4 18.6 0.896 1.511 7.7 0.099 -0.21
World-R1 18.8 18.1 19.4 0.899 1.550 7.7 0.101 -0.20
Global-only (ours) 19.4 18.6 19.7 0.900 1.396 7.6 0.099 -0.21
LoGo (ours) 19.7 19.0 20.0 0.905 1.352 7.5 0.101 -0.19
Complete original PDF block S4.T2
Table 2: Results on DL3DV. LoGo yields larger gains than baseline methods across three base models for all 3D consistency metrics, while maintaining camera control and video quality, with up to 37%37\% decrease in epipolar error. Standard errors are shown in Table 7 in the Appendix.
Original selectable table
Model 3D consistency Camera Video quality
PSNR-V ↑\uparrow PSNR-D ↑\uparrow PSNR-GS ↑\uparrow MVCS ↑\uparrow Epipolar ↓\downarrow RPE∘ ↓\downarrow RPE trans ↓\downarrow VR-VQ ↑\uparrow
Lingbot2 Base 13.3 13.1 15.6 0.719 2.161 20.52 0.082 -0.422
VideoGPA 13.3 13.2 15.6 0.728 2.429 20.15 0.081 -0.409
World-R1 13.6 13.3 15.7 0.717 2.271 20.13 0.070 -0.429
Global-only (ours) 14.0 13.8 16.3 0.772 2.107 22.68 0.086 -0.305
LoGo (ours) 16.1 15.8 17.3 0.795 1.365 18.33 0.077 -0.283
UniWorld Base 14.7 14.5 16.6 0.748 1.510 12.94 0.057 -0.128
VideoGPA 14.7 14.6 16.7 0.752 1.623 13.31 0.058 -0.129
World-R1 14.9 14.7 17.3 0.747 1.926 15.45 0.064 -0.234
Global-only (ours) 14.9 14.7 16.7 0.778 1.557 12.18 0.058 -0.134
LoGo (ours) 15.4 15.1 17.4 0.807 1.429 12.64 0.055 -0.121
Lyra2 Base 15.2 14.8 16.8 0.819 2.448 22.72 0.148 -0.557
VideoGPA 15.0 14.7 16.6 0.813 2.595 21.95 0.143 -0.508
World-R1 15.3 15.2 16.9 0.832 2.213 21.72 0.140 -0.523
Global-only (ours) 15.6 15.3 17.1 0.826 2.244 22.17 0.145 -0.525
LoGo (ours) 15.9 15.6 17.3 0.844 1.863 22.38 0.145 -0.513
Complete original PDF block S4.F7
Figure 6: Qualitative results of LoGo vs. baselines. Red boxes indicate inconsistent regions. LoGo effectively fixes both global and local inconsistencies, including scene change (top right), hallucinated objects (top left), floater artifacts (middle), and inconsistent object appearance (bottom). Baseline post-training methods are not capable of resolving these issues, while LoGo is.
Complete original PDF block S4.F7b
Figure 7: Global-only reward vs. LoGo. A 3D-consistent video should closely match its reprojections, though not perfectly, since keyframes in long videos overlap only sparsely. Global reward incurs inconsistencies, highlighted in red, such as hallucinated object (double railing), artifact (floater in the robot scene), and object appearance change (chair). LoGo, by combining global reward with local reward, effectively removes these inconsistencies.
Complete original PDF block S4.T3
Table 3: Difficulty breakdown on TrajectoryBench. We show results on easy, medium, and hard categories of TrajectoryBench. PSNR-V denotes RGBD reprojection error with VGGT, and PSNR-D with DA3. LoGo’s margin over the baselines widens as difficulty increases across all three base models. This trend also holds for other metrics found in Table 6 in the Appendix.
Original selectable table
Model PSNR-V ↑\uparrow PSNR-D ↑\uparrow Epipolar ↓\downarrow
E M H E M H E M H
Lingbot2 Base 22.6 18.2 15.3 20.6 17.6 14.5 0.524 0.681 2.034
VideoGPA 22.7 18.3 15.4 20.7 17.7 14.6 0.533 0.684 1.997
World-R1 22.7 18.1 15.4 20.6 17.4 14.5 0.526 0.681 2.046
LoGo 23.8 20.2 17.6 21.6 19.7 17.0 0.490 0.601 1.439
UniWorld Base 23.0 17.2 16.7 21.2 16.4 16.1 0.450 0.760 1.623
VideoGPA 23.1 17.2 16.8 21.2 16.4 16.2 0.450 0.773 1.642
World-R1 22.4 17.3 17.2 20.5 16.5 16.6 0.462 0.860 1.805
LoGo 23.2 18.2 18.1 21.3 17.4 17.5 0.446 0.760 1.379
Lyra2 Base 22.2 17.9 17.8 21.1 17.2 17.3 0.519 0.874 2.027
VideoGPA 22.1 17.7 17.3 21.1 17.1 16.7 0.519 0.878 1.963
World-R1 22.2 18.1 18.4 21.0 17.5 17.8 0.529 0.939 1.999
LoGo 22.5 18.8 19.4 21.2 18.1 18.8 0.510 0.847 1.723

5 Ablations and Analyses

For Lingbot2, the reported epipolar-error reduction grows from 24% at 80 frames to 62% at 400 frames. This extends the test horizon beyond the main 240-frame benchmark but does not establish reliable generation beyond 400 frames.

Voxel credit beats 40-frame windows, framewise and 2D-patch alternatives on the hard-split geometry columns. Table 4’s geometry values are the hard-split values from Table 8, while its camera and quality columns are the aggregate values. Voxel feedback does not win every camera column: window-based translation error is .094 versus voxel .107.

Tables 5 and 11 expose the tradeoff. Local-only training raises geometry scores but lowers HPSv3 from 5.09 to 4.47. Interleaving restores it to 4.94; adding the global component reaches 5.18 and improves translation error. Some geometry measures peak at the interleaved or local-only stages rather than the final recipe, so “best overall” reflects a balance of objectives.

The appendix tests other optimizers and voxel sizes. Improvements often persist, but not in every cell: under Flow-GRPO the hard epipolar error worsens from global-only 1.500 to local/global 1.553. These ablations support useful transfer across algorithms rather than a universal dominance claim.

Complete original PDF block S5.F8
Figure 8: Epipolar error vs. length.
Complete original PDF block S5.T4
Table 4: Comparison of different local rewards. We compare voxel (used in LoGo) with framewise, 2D patchwise, and 40-frame window credit assignment. Voxel-based yields the best consistency gains. Easy/medium/hard breakdown and standard errors are in Table 8 in the Appendix.
Original selectable table
3D Consistency Camera Video Quality
PSNR-V ↑\uparrow PSNR-D ↑\uparrow PSNR-GS ↑\uparrow MVCS ↑\uparrow Epipolar ↓\downarrow RPE∘ ↓\downarrow RPE trans ↓\downarrow VR-VQ ↑\uparrow
Base 15.3 14.5 15.7 0.812 2.034 13.2 0.108 0.21
Global-only 16.5 15.7 16.7 0.848 1.726 11.2 0.102 0.31
Window 17.0 16.3 17.1 0.856 1.666 9.5 0.094 0.32
Framewise 17.0 16.4 16.9 0.857 1.531 10.7 0.101 0.24
Patchwise 16.9 16.2 17.0 0.865 1.660 8.4 0.100 0.24
Voxel 17.8 17.2 17.7 0.874 1.381 8.9 0.107 0.38
Complete original PDF block S5.T5
Table 5: Ablating the LoGo recipe. Using local-only reward yields best 3D consistency but drops camera and aesthetics metrics. Adding in reward interleaving improves other metrics, but still degrades HPSv3 compared to the base model. LoGo, which includes both reward interleaving and local-global blending, yields the best overall result. Standard errors are in Table 11 in the Appendix.
Original selectable table
3D Consistency Camera Video Quality
PSNR-V ↑\uparrow PSNR-D ↑\uparrow PSNR-GS ↑\uparrow MVCS ↑\uparrow Epipolar ↓\downarrow RPE∘ ↓\downarrow RPE trans ↓\downarrow VR-VQ ↑\uparrow HPSv3 ↑\uparrow
Base 16.9 16.0 16.9 0.842 1.507 13.2 0.108 0.21 5.09
Local-only 19.2 18.2 18.7 0.893 1.078 11.2 0.116 0.32 4.47
+ interleave 19.2 18.3 18.6 0.897 1.072 8.9 0.107 0.38 4.94
+ global (final) 19.0 18.2 18.4 0.894 1.111 8.9 0.103 0.37 5.18

6 Conclusion and Limitations

LoGo provides a practical reward design for static, camera-controlled scene generation and a benchmark that reveals long-path consistency problems. It is not a new guarantee of a physically persistent world. Reconstruction bias, reward exploitation and incomplete implementation details remain relevant when reproducing it.

The authors explicitly leave dynamic scenes and horizons beyond 400 frames unresolved. They suggest stronger memory for longer videos and dynamic 4D reconstruction for moving scenes; neither extension is demonstrated here.

7 AI use statement

The authors disclose using language models for sentence polishing, schematic illustrations and routine implementation assistance. They state that generative tools did not supply the core research idea or paper content.

8 Acknowledgments

The work acknowledges the World Labs team for discussion and feedback.

Appendix A Appendix

The supplement supplies the patch-level training objective, six detailed numerical tables, training and evaluation settings, benchmark construction and additional reward/voxel visualizations. All remain available below.

A.1 RL reward formulation with LoGo

The original DiffusionNFT objective mixes positive and negative velocity-target losses with one sample-level optimality probability. Equation 3 averages voxel probabilities over the set of voxels whose projected pixels intersect a latent space-time patch, then applies that patch-specific weighting to its two squared-error terms.

This is an unweighted average over intersecting voxels as written, not an average proportional to each voxel’s pixel count. It provides different credit to different patches within the same video. Empty-intersection handling and other numerical details are not specified here; the source states that Flow-GRPO can be decomposed similarly.

Equations 3
ℒNFT=𝔼⁡[r​‖vθ+−v‖2+(1−r)​‖vθ−−v‖2]\displaystyle\mathcal{L}_{\mathrm{NFT}}=\mathbb{E}\Big[r\|v_{\theta}^{+}-v\|^{2}+(1-r)\|v_{\theta}^{-}-v\|^{2}\Big]
ℒNFT,j(k)=1|A|𝔼[∑i∈A={i:Si∩Ij​(k)≠∅}ri∥vθ+,j(k)−vj(k)∥2+(1−ri)∥vθ−,j(k)−vj(k)∥2]\displaystyle\mathcal{L}_{\mathrm{NFT},j}^{(k)}=\frac{1}{|A|}\mathbb{E}\Big[\sum_{i\in A=\{i:S_{i}\cap I_{j}(k)\neq\emptyset\}}r_{i}\|v_{\theta^{+},j}^{(k)}-v_{j}^{(k)}\|^{2}+(1-r_{i})\|v_{\theta^{-},j}^{(k)}-v_{j}^{(k)}\|^{2}\Big] (3)

A.2 Additional experiment results

Tables 6 and 7 retain all model/baseline means and paired standard errors. The errors are reported as paired evaluation errors; the text does not give a repeated-training-seed study or a universal significance test for every comparison. Small differences and ties should be read at their displayed precision.

Table 8 expands the localization comparison: its geometry columns split easy/medium/hard, while camera and VQ remain aggregate. Voxel geometry is usually strongest, but some easy epipolar or tied PSNR cells favor another readout. Its voxel row is not numerically identical to the final main-recipe row and should be read as a separate ablation.

Table 9 compares global-only versus local/global rewards under three optimizers on UniWorld. The paired cells show mostly favorable reconstruction changes with exceptions in epipolar, camera, VQ and some tied geometry entries. The DiffusionNFT ablation values also differ from the main UniWorld row; the paper does not establish that all are the same checkpoint.

Table 10 gives voxel sizes .05, .1, .3, .5 and .7. The caption says median scene depth, whereas the method and Figure 10 use the 90th percentile; this discrepancy is preserved. The .1 choice is a useful operating point, not the maximum of every column. Table 11 retains the full geometry/quality tradeoff and uncertainty for the recipe components.

Complete original PDF block A1.T6
Table 6: TrajectoryBench evaluation difficulty breakdown. Geometry columns are split by difficulty tier E / M / H for easy, medium, hard splits. LoGo achieves better 3D consistency than baseline methods in all twelve metrics across three splits for all three models, while maintaining or improving camera control and video quality, with the strongest advantage on the hard split. ±\pm denotes paired standard error.
Original selectable table
Model 3D consistency Camera Video quality
PSNR-V ↑\uparrow PSNR-D ↑\uparrow PSNR-GS ↑\uparrow MVCS ↑\uparrow Epipolar ↓\downarrow RPE∘ ↓\downarrow RPE trans ↓\downarrow VR-VQ ↑\uparrow
E M H E M H E M H E M H E M H
Lingbot2 Base 22.6 18.2 15.3 20.6 17.6 14.5 21.0 18.0 15.7 0.918 0.878 0.812 0.524 0.681 2.034 13.2 0.108 0.21
VideoGPA 22.7±\pm0.03 18.3±\pm0.08 15.4±\pm0.04 20.7±\pm0.05 17.7±\pm0.08 14.6±\pm0.05 21.1±\pm0.05 18.1±\pm0.07 15.8±\pm0.04 0.916±\pm0.002 0.880±\pm0.004 0.820±\pm0.005 0.533±\pm0.011 0.684±\pm0.013 1.997±\pm0.052 13.1±\pm0.14 0.107±\pm0.001 0.22±\pm0.004
World-R1 22.7±\pm0.06 18.1±\pm0.08 15.4±\pm0.04 20.6±\pm0.08 17.4±\pm0.08 14.5±\pm0.05 21.1±\pm0.06 18.0±\pm0.07 15.7±\pm0.04 0.917±\pm0.002 0.873±\pm0.005 0.809±\pm0.005 0.526±\pm0.012 0.681±\pm0.012 2.046±\pm0.054 13.1±\pm0.14 0.108±\pm0.001 0.21±\pm0.004
Global (ours) 22.9±\pm0.14 19.3±\pm0.11 16.5±\pm0.05 20.8±\pm0.13 18.8±\pm0.11 15.7±\pm0.06 21.4±\pm0.11 18.9±\pm0.09 16.7±\pm0.05 0.916±\pm0.005 0.908±\pm0.006 0.848±\pm0.005 0.521±\pm0.013 0.649±\pm0.015 1.726±\pm0.057 11.2±\pm0.17 0.102±\pm0.001 0.31±\pm0.006
LoGo (ours) 23.8±\pm0.16 20.2±\pm0.11 17.6±\pm0.05 21.6±\pm0.15 19.7±\pm0.11 17.0±\pm0.06 21.7±\pm0.11 19.4±\pm0.09 17.4±\pm0.05 0.942±\pm0.006 0.929±\pm0.006 0.871±\pm0.006 0.490±\pm0.010 0.601±\pm0.012 1.439±\pm0.052 8.9±\pm0.18 0.103±\pm0.001 0.37±\pm0.006
UniWorld Base 23.0 17.2 16.7 21.2 16.4 16.1 20.6 17.5 16.9 0.900 0.898 0.871 0.450 0.760 1.623 13.8 0.132 0.39
VideoGPA 23.1±\pm0.02 17.2±\pm0.02 16.8±\pm0.02 21.2±\pm0.05 16.4±\pm0.04 16.2±\pm0.03 20.6±\pm0.05 17.6±\pm0.03 16.9±\pm0.02 0.900±\pm0.001 0.898±\pm0.001 0.871±\pm0.002 0.450±\pm0.002 0.773±\pm0.010 1.642±\pm0.036 13.8±\pm0.06 0.132±\pm0.000 0.39±\pm0.002
World-R1 22.4±\pm0.07 17.3±\pm0.08 17.2±\pm0.06 20.5±\pm0.08 16.5±\pm0.08 16.6±\pm0.06 20.1±\pm0.09 17.9±\pm0.08 18.0±\pm0.05 0.894±\pm0.003 0.883±\pm0.006 0.858±\pm0.005 0.462±\pm0.003 0.860±\pm0.015 1.805±\pm0.042 13.1±\pm0.19 0.129±\pm0.001 0.22±\pm0.006
Global (ours) 22.8±\pm0.05 17.6±\pm0.05 17.6±\pm0.05 20.8±\pm0.06 16.9±\pm0.07 17.1±\pm0.06 20.5±\pm0.07 17.9±\pm0.06 17.6±\pm0.04 0.899±\pm0.001 0.901±\pm0.005 0.887±\pm0.003 0.447±\pm0.003 0.773±\pm0.017 1.514±\pm0.039 13.3±\pm0.14 0.126±\pm0.001 0.37±\pm0.004
LoGo (ours) 23.2±\pm0.05 18.2±\pm0.06 18.1±\pm0.06 21.3±\pm0.07 17.4±\pm0.07 17.5±\pm0.06 21.0±\pm0.09 18.5±\pm0.06 18.2±\pm0.05 0.904±\pm0.002 0.913±\pm0.004 0.895±\pm0.003 0.446±\pm0.002 0.760±\pm0.012 1.379±\pm0.036 14.0±\pm0.18 0.130±\pm0.001 0.41±\pm0.004
Lyra2 Base 22.2 17.9 17.8 21.1 17.2 17.3 21.6 18.9 18.5 0.924 0.922 0.881 0.519 0.874 2.027 7.9 0.100 -0.23
VideoGPA 22.1±\pm0.04 17.7±\pm0.06 17.3±\pm0.04 21.1±\pm0.05 17.1±\pm0.06 16.7±\pm0.04 21.6±\pm0.06 18.7±\pm0.06 18.0±\pm0.04 0.920±\pm0.003 0.926±\pm0.003 0.880±\pm0.003 0.519±\pm0.003 0.878±\pm0.033 1.963±\pm0.050 7.7±\pm0.11 0.099±\pm0.001 -0.21±\pm0.003
World-R1 22.2±\pm0.05 18.1±\pm0.06 18.4±\pm0.04 21.0±\pm0.07 17.5±\pm0.06 17.8±\pm0.04 21.7±\pm0.07 19.3±\pm0.08 19.0±\pm0.04 0.920±\pm0.003 0.929±\pm0.003 0.883±\pm0.003 0.529±\pm0.007 0.939±\pm0.034 1.999±\pm0.051 7.7±\pm0.12 0.101±\pm0.001 -0.20±\pm0.003
Global (ours) 22.5±\pm0.05 18.6±\pm0.06 19.0±\pm0.04 21.3±\pm0.07 17.9±\pm0.06 18.4±\pm0.04 21.8±\pm0.07 19.6±\pm0.07 19.4±\pm0.04 0.921±\pm0.003 0.929±\pm0.003 0.884±\pm0.003 0.513±\pm0.003 0.869±\pm0.023 1.784±\pm0.050 7.6±\pm0.11 0.099±\pm0.001 -0.21±\pm0.003
LoGo (ours) 22.5±\pm0.04 18.8±\pm0.06 19.4±\pm0.04 21.2±\pm0.06 18.1±\pm0.07 18.8±\pm0.05 21.8±\pm0.08 19.8±\pm0.07 19.7±\pm0.04 0.922±\pm0.003 0.930±\pm0.003 0.892±\pm0.003 0.510±\pm0.003 0.847±\pm0.025 1.723±\pm0.051 7.5±\pm0.10 0.101±\pm0.001 -0.19±\pm0.003
Complete original PDF block A1.T7
Table 7: DL3DV evaluation with standard errors. ±\pm denotes paired standard errors. LoGo yields more improvement in 3D consistency compared to baseline post-training methods on three different base models for all four metrics, while maintaining camera control and video quality. The improvement is significant, with up to 37%37\% decrease in epipolar error.
Original selectable table
Model 3D consistency Camera Video quality
PSNR-V ↑\uparrow PSNR-D ↑\uparrow PSNR-GS ↑\uparrow MVCS ↑\uparrow Epipolar ↓\downarrow RPE∘ ↓\downarrow RPE trans ↓\downarrow VR-VQ ↑\uparrow
Lingbot2 Base 13.3 13.1 15.6 0.719 2.161 20.52 0.082 -0.422
VideoGPA 13.3±\pm0.10 13.2±\pm0.11 15.6±\pm0.09 0.728±\pm0.014 2.429±\pm0.158 20.15±\pm1.215 0.081±\pm0.005 -0.409±\pm0.010
World-R1 13.6±\pm0.11 13.3±\pm0.10 15.7±\pm0.11 0.717±\pm0.017 2.271±\pm0.154 20.13±\pm1.681 0.070±\pm0.004 -0.429±\pm0.012
Global (ours) 14.0±\pm0.15 13.8±\pm0.14 16.3±\pm0.14 0.772±\pm0.016 2.107±\pm0.152 22.68±\pm1.745 0.086±\pm0.006 -0.305±\pm0.017
LoGo (ours) 16.1±\pm0.22 15.8±\pm0.20 17.3±\pm0.21 0.795±\pm0.021 1.365±\pm0.198 18.33±\pm1.504 0.077±\pm0.006 -0.283±\pm0.018
UniWorld Base 14.7 14.5 16.6 0.748 1.510 12.94 0.057 -0.128
VideoGPA 14.7±\pm0.03 14.6±\pm0.07 16.7±\pm0.05 0.752±\pm0.005 1.623±\pm0.082 13.31±\pm0.458 0.058±\pm0.001 -0.129±\pm0.008
World-R1 14.9±\pm0.12 14.7±\pm0.12 17.3±\pm0.14 0.747±\pm0.018 1.926±\pm0.102 15.45±\pm1.432 0.064±\pm0.002 -0.234±\pm0.015
Global (ours) 14.9±\pm0.10 14.7±\pm0.11 16.7±\pm0.07 0.778±\pm0.013 1.557±\pm0.079 12.18±\pm0.514 0.058±\pm0.002 -0.134±\pm0.011
LoGo (ours) 15.4±\pm0.11 15.1±\pm0.12 17.4±\pm0.09 0.807±\pm0.014 1.429±\pm0.072 12.64±\pm0.949 0.055±\pm0.002 -0.121±\pm0.012
Lyra2 Base 15.2 14.8 16.8 0.819 2.448 22.72 0.148 -0.557
VideoGPA 15.0±\pm0.09 14.7±\pm0.08 16.6±\pm0.09 0.813±\pm0.011 2.595±\pm0.216 21.95±\pm0.549 0.143±\pm0.003 -0.508±\pm0.011
World-R1 15.3±\pm0.11 15.2±\pm0.13 16.9±\pm0.12 0.832±\pm0.010 2.213±\pm0.200 21.72±\pm0.768 0.140±\pm0.003 -0.523±\pm0.011
Global (ours) 15.6±\pm0.11 15.3±\pm0.13 17.1±\pm0.10 0.826±\pm0.013 2.244±\pm0.253 22.17±\pm0.619 0.145±\pm0.003 -0.525±\pm0.010
LoGo (ours) 15.9±\pm0.12 15.6±\pm0.13 17.3±\pm0.10 0.844±\pm0.012 1.863±\pm0.225 22.38±\pm0.690 0.145±\pm0.002 -0.513±\pm0.012
Complete original PDF block A1.T8
Table 8: Full ablation table of different ways to localize the reward. We compare voxel reward (used in LoGo as local reward), and other localization schemes such as framewise, 2D patchwise, frame windows (40f windows). Voxel localization yields the best improvement because it normalizes per-group rollouts in the native 3D space where 3D consistency is measured, without being affected by shifts in region represented per frame/patch caused by slightly different camera following in each rollout. This advantage persists across difficulty splits, and is most significant on the “hard” split, which has longer generation horizon and more complex camera motions.
Original selectable table
Reward readout 3D consistency Camera Video quality
PSNR-V ↑\uparrow PSNR-D ↑\uparrow PSNR-GS ↑\uparrow MVCS ↑\uparrow Epipolar ↓\downarrow RPE∘ ↓\downarrow RPE trans ↓\downarrow VR-VQ ↑\uparrow
E M H E M H E M H E M H E M H
Base 22.6 18.2 15.3 20.6 17.6 14.5 21.0 18.0 15.7 0.918 0.878 0.812 0.524 0.681 2.034 13.2 0.108 0.21
Global 22.9±\pm0.14 19.3±\pm0.11 16.5±\pm0.05 20.8±\pm0.13 18.8±\pm0.11 15.7±\pm0.06 21.4±\pm0.11 18.9±\pm0.09 16.7±\pm0.05 0.916±\pm0.005 0.908±\pm0.006 0.848±\pm0.005 0.521±\pm0.013 0.649±\pm0.015 1.726±\pm0.057 11.2±\pm0.17 0.102±\pm0.001 0.31±\pm0.006
Window 23.3±\pm0.17 19.9±\pm0.11 17.0±\pm0.05 21.2±\pm0.15 19.3±\pm0.11 16.3±\pm0.06 21.6±\pm0.12 19.4±\pm0.09 17.1±\pm0.05 0.918±\pm0.005 0.914±\pm0.006 0.856±\pm0.005 0.539±\pm0.020 0.635±\pm0.014 1.666±\pm0.053 9.5±\pm0.17 0.094±\pm0.001 0.32±\pm0.006
Framewise 23.5±\pm0.17 19.7±\pm0.11 17.0±\pm0.05 21.3±\pm0.16 19.0±\pm0.11 16.4±\pm0.06 21.5±\pm0.11 19.0±\pm0.09 16.9±\pm0.05 0.929±\pm0.005 0.918±\pm0.006 0.857±\pm0.005 0.531±\pm0.021 0.617±\pm0.013 1.531±\pm0.053 10.7±\pm0.17 0.101±\pm0.001 0.24±\pm0.006
Patchwise 23.4±\pm0.16 19.6±\pm0.11 16.9±\pm0.06 21.2±\pm0.15 18.8±\pm0.11 16.2±\pm0.06 21.5±\pm0.12 19.0±\pm0.09 17.0±\pm0.05 0.930±\pm0.005 0.918±\pm0.006 0.865±\pm0.005 0.509±\pm0.010 0.621±\pm0.013 1.660±\pm0.054 8.4±\pm0.18 0.100±\pm0.001 0.24±\pm0.006
Voxel 23.9±\pm0.14 20.3±\pm0.11 17.8±\pm0.05 21.6±\pm0.14 19.7±\pm0.11 17.2±\pm0.06 21.7±\pm0.11 19.4±\pm0.09 17.7±\pm0.05 0.937±\pm0.005 0.934±\pm0.006 0.874±\pm0.005 0.499±\pm0.029 0.587±\pm0.013 1.381±\pm0.053 8.9±\pm0.18 0.107±\pm0.001 0.38±\pm0.006
Complete original PDF block A1.T9
Table 9: Full difficulty breakdown of TrajectoryBench evaluation of LoGo with three different post-training algorithms: DiffusionNFT, Flow-GRPO, and DRaFT. LoGo’s advantage persists regardless of post-training algorithms. Paired standard error is displayed in each cell.
Original selectable table
RL algorithm 3D consistency Camera Video quality
PSNR-V ↑\uparrow PSNR-D ↑\uparrow PSNR-GS ↑\uparrow MVCS ↑\uparrow Epipolar ↓\downarrow RPE∘ ↓\downarrow RPE trans ↓\downarrow VR-VQ ↑\uparrow
E M H E M H E M H E M H E M H
Base 23.0 17.2 16.7 21.2 16.4 16.1 20.6 17.5 16.9 0.900 0.898 0.871 0.450 0.760 1.623 13.8 0.132 0.39
DiffusionNFT Ours-global 22.8±\pm0.05 17.6±\pm0.05 17.6±\pm0.05 20.8±\pm0.06 16.9±\pm0.07 17.1±\pm0.06 20.5±\pm0.07 17.9±\pm0.06 17.6±\pm0.04 0.899±\pm0.001 0.901±\pm0.005 0.887±\pm0.003 0.447±\pm0.003 0.773±\pm0.017 1.514±\pm0.039 13.3±\pm0.14 0.126±\pm0.001 0.37±\pm0.004
LoGlo 23.1±\pm0.05 17.9±\pm0.06 17.9±\pm0.06 21.2±\pm0.07 17.2±\pm0.07 17.4±\pm0.06 20.9±\pm0.08 18.4±\pm0.07 18.0±\pm0.05 0.903±\pm0.002 0.912±\pm0.004 0.895±\pm0.003 0.446±\pm0.003 0.781±\pm0.017 1.436±\pm0.037 13.9±\pm0.18 0.130±\pm0.001 0.40±\pm0.004
FlowGRPO Ours-global 23.2±\pm0.03 17.7±\pm0.03 17.5±\pm0.03 21.3±\pm0.06 16.8±\pm0.04 16.8±\pm0.04 20.9±\pm0.10 18.0±\pm0.04 17.5±\pm0.03 0.901±\pm0.001 0.902±\pm0.002 0.880±\pm0.002 0.456±\pm0.001 0.788±\pm0.014 1.500±\pm0.035 13.8±\pm0.07 0.132±\pm0.000 0.37±\pm0.003
LoGlo 23.3±\pm0.03 17.8±\pm0.03 17.6±\pm0.03 21.5±\pm0.05 17.0±\pm0.04 16.9±\pm0.04 21.0±\pm0.08 18.1±\pm0.04 17.6±\pm0.03 0.902±\pm0.001 0.902±\pm0.002 0.883±\pm0.002 0.454±\pm0.002 0.764±\pm0.008 1.553±\pm0.039 13.7±\pm0.10 0.132±\pm0.000 0.36±\pm0.003
DRaFT Ours-global 23.0±\pm0.02 17.1±\pm0.02 16.6±\pm0.02 21.1±\pm0.05 16.2±\pm0.04 16.0±\pm0.03 20.6±\pm0.07 17.4±\pm0.03 16.7±\pm0.02 0.899±\pm0.000 0.894±\pm0.001 0.869±\pm0.002 0.453±\pm0.004 0.768±\pm0.011 1.726±\pm0.037 14.0±\pm0.07 0.132±\pm0.000 0.39±\pm0.003
LoGlo 23.0±\pm0.02 17.2±\pm0.02 16.8±\pm0.02 21.1±\pm0.04 16.3±\pm0.04 16.2±\pm0.03 20.6±\pm0.06 17.5±\pm0.03 16.8±\pm0.02 0.899±\pm0.001 0.893±\pm0.002 0.871±\pm0.002 0.449±\pm0.002 0.765±\pm0.013 1.670±\pm0.038 14.0±\pm0.08 0.132±\pm0.000 0.39±\pm0.003
Complete original PDF block A1.T10
Table 10: Ablation of localization voxel size on TrajectoryBench, with difficulty breakdown and paired standard errors. Decrease of voxel size generally increases 3D consistency, up to 0.1 ×\times median scene depth. We choose voxel size of 0.1 based on this ablation.
Original selectable table
Voxel size 3D consistency Camera Video quality
PSNR-V ↑\uparrow PSNR-D ↑\uparrow PSNR-GS ↑\uparrow MVCS ↑\uparrow Epipolar ↓\downarrow RPE∘ ↓\downarrow RPE trans ↓\downarrow VR-VQ ↑\uparrow
E M H E M H E M H E M H E M H
Base 22.6 18.2 15.3 20.6 17.6 14.5 21.0 18.0 15.7 0.918 0.878 0.812 0.524 0.681 2.034 13.2 0.108 0.21
0.05 24.1±\pm0.16 20.7±\pm0.11 18.6±\pm0.06 21.8±\pm0.16 20.1±\pm0.11 17.8±\pm0.06 21.7±\pm0.13 19.7±\pm0.09 18.2±\pm0.06 0.932±\pm0.005 0.935±\pm0.007 0.880±\pm0.005 0.466±\pm0.010 0.579±\pm0.014 1.229±\pm0.051 9.2±\pm0.17 0.108±\pm0.001 0.42±\pm0.006
0.1 24.3±\pm0.15 20.5±\pm0.11 18.7±\pm0.06 21.8±\pm0.15 19.9±\pm0.11 17.9±\pm0.07 21.8±\pm0.10 19.6±\pm0.09 18.4±\pm0.06 0.941±\pm0.005 0.939±\pm0.007 0.885±\pm0.006 0.486±\pm0.014 0.577±\pm0.014 1.235±\pm0.051 9.0±\pm0.19 0.111±\pm0.001 0.42±\pm0.006
0.3 24.6±\pm0.16 20.7±\pm0.11 18.4±\pm0.06 22.2±\pm0.16 20.2±\pm0.11 17.6±\pm0.07 21.7±\pm0.11 19.6±\pm0.09 18.0±\pm0.06 0.947±\pm0.006 0.940±\pm0.007 0.893±\pm0.006 0.467±\pm0.012 0.581±\pm0.013 1.246±\pm0.051 10.5±\pm0.18 0.117±\pm0.001 0.42±\pm0.006
0.5 23.6±\pm0.12 19.9±\pm0.11 17.5±\pm0.06 21.4±\pm0.12 19.3±\pm0.11 16.6±\pm0.06 21.6±\pm0.10 19.2±\pm0.09 17.5±\pm0.05 0.933±\pm0.004 0.923±\pm0.006 0.865±\pm0.005 0.484±\pm0.016 0.595±\pm0.013 1.450±\pm0.053 14.5±\pm0.18 0.119±\pm0.001 0.39±\pm0.006
0.7 24.2±\pm0.17 20.3±\pm0.11 17.7±\pm0.06 21.8±\pm0.16 19.7±\pm0.11 17.0±\pm0.06 21.8±\pm0.11 19.4±\pm0.09 17.6±\pm0.05 0.935±\pm0.005 0.929±\pm0.006 0.882±\pm0.005 0.489±\pm0.014 0.588±\pm0.013 1.321±\pm0.052 10.3±\pm0.19 0.107±\pm0.001 0.43±\pm0.006
Complete original PDF block A1.T11
Table 11: Ablation on LoGo recipe with paired standard errors. Using local reward only yields best 3D consistency but worst camera and aesthetics metrics. Adding in reward interleaving improves other metrics, but still degrades HPSv3 compared to the base model. LoGo, which includes both reward interleaving and local-global blending, yields the best overall result.
Original selectable table
3D consistency Camera Video quality
PSNR-V ↑\uparrow PSNR-D ↑\uparrow PSNR-GS ↑\uparrow MVCS ↑\uparrow Epipolar ↓\downarrow RPE∘ ↓\downarrow RPE trans ↓\downarrow VR-VQ ↑\uparrow HPSv3 ↑\uparrow
Base 16.9 16.0 16.9 0.842 1.507 13.2 0.108 0.21 5.09
Local reward 19.2±\pm0.05 18.2±\pm0.05 18.7±\pm0.04 0.893±\pm0.004 1.078±\pm0.034 11.2±\pm0.12 0.116±\pm0.001 0.32±\pm0.004 4.47±\pm0.019
+ interleave 19.2±\pm0.05 18.3±\pm0.05 18.6±\pm0.04 0.897±\pm0.004 1.072±\pm0.034 8.9±\pm0.13 0.107±\pm0.001 0.38±\pm0.004 4.94±\pm0.013
+ global (final) 19.0±\pm0.05 18.2±\pm0.05 18.4±\pm0.04 0.894±\pm0.004 1.111±\pm0.033 8.9±\pm0.13 0.103±\pm0.001 0.37±\pm0.004 5.18±\pm0.010

A.3 Additional training details

The first half of each training camera trajectory is subsampled to 240 frames. COLMAP translations are scaled by .4 for Lyra2 and 1 for UniWorld; Lingbot2 normalizes by each scene’s largest step. Batch size is eight and rollout-group size sixteen. Learning rates are 3×10⁻⁴ for Lingbot2/UniWorld and 3.33×10⁻⁵ for Lyra2.

RGB/depth reward weights are .5/.5 for Lingbot2 and UniWorld, and .3/.7 for Lyra2 to reduce reward exploitation. UniWorld uses BLIP-2 captions during both training and inference; the other two models do not use captions. Training stops around reward convergence, taking 16–36 hours on 64 H100s. Exact update counts, schedule ratios and training-seed uncertainty are not fully reported.

A.4 Additional evaluation details

Reprojection evaluation fuses keyframes sampled every twentieth frame while excluding the frame being scored. This leave-one-out construction prevents a frame’s own points from dominating its reprojection score when distant keyframes have sparse spatial overlap. Sparse coverage and estimated depth/poses still limit interpretation.

A.5 Additional details on TrajectoryBench construction

Easy tasks reuse a WorldScore subset. Medium tasks combine three motions in one generation rather than WorldScore’s separate single-motion clips, and add doorway transitions. Hard indoor/outdoor initial images come from Pexels; doorway images use Nano Banana to control door placement while varying scene types.

Human-designed motion primitives are mapped to paths and tried on base generators to adjust translation and rotation so views cover substantial space without apparent wall/object collisions. The benchmark is therefore deliberately constructed and tuned, not an independently sampled collision-certified trajectory distribution.

Twenty percent of tasks are stylized with Nano Banana into eight visual styles, from anime and oil painting to low-poly, cyberpunk and steampunk. The paper lists the scene/style coverage but does not give every split count, random seed or a formal scene-overlap audit against foundation-model training.

A.6 Additional Analyses and Ablations

The final analyses show how voxel-local feedback highlights a transient object and how the spatial grouping scale affects reconstruction and camera behavior.

A.6.1 Additional visualization of LoGo’s local reward.

Figure 9 projects normalized voxel scores back to image patches. Red feedback concentrates on a hallucinated tree that later splits into two; the LoGo example removes it. This is an illustrative localized correction, not a frequency estimate over all possible hallucinations.

Complete original PDF block A1.F9
Figure 9: Visualization of LoGo’s local reward. The red-blue overlay is the normalized score per voxel projected back to image patches - blue means good and red means bad. In this example, a tree that did not exist in earlier is hallucinated in the middle of the video and splits into two. This is caught by LoGo’s local reward component, which concentrates on the tree region, and subsequently fixed by post-training with LoGo.

A.6.2 Voxel granularity ablation.

Figure 10 and Table 10 test coarser and finer voxels around the .1×depth operating point. Finer grouping can isolate an inconsistent region more precisely, but results are not monotonic across all metrics or difficulty splits. The .05 setting already beats .1 on some geometry cells, while .3 is stronger on some easy/medium measures.

The proposed camera explanation is that a coarse penalty can encourage turning away from a large region, whereas a small voxel makes avoidance less attractive. This is a hypothesis about the optimization incentive, not an isolated causal experiment. Use the body’s stated P90 convention while retaining the conflicting median-depth table caption.

Complete original PDF block A1.F10
Figure 10: Effect of voxel size on geometry metric and camera following metrics. Both improve with finer voxel size up till 0.1×\times P90 scene depth.
Concept demonstration

Open image ↗