cs.CVOct 2, 2026

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

Authors: Ziqi Ma, Shreya Sharma, Mohamed El Banani, Katja Schwarz, Chongjie Ye, Chao-Yuan Wu, Li Fei-Fei, Ben Mildenhall, 

Organizations: California Institute of Technology · World Labs · Work done during an internship at World Labs

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Camera-controlled video models are rapidly advancing toward long generation horizons and complex camera control. A key failure mode is 3D inconsistency: as the camera moves, objects lose permanence and scene structures shift. Existing post-training techniques, which assign a single scalar reward to the entire generation, are poorly suited to correcting these inconsistencies over long horizons. We introduce LoGo, which blends global and spatially localized rewards for camera-controlled video models. The local reward provides fine-grained credit assignment, which substantially improves 3D consistency, while the global reward preserves camera following and video quality. Across three base models, LoGo shows a clear advantage on DL3DV and TrajectoryBench, a new benchmark for long-horizon, complex-camera-control generation that current evaluations lack. LoGo effectively reduces local object shifts, artifacts, and global scene changes, illustrating the importance of credit assignment in post-training video models. Project website: https://ziqi-ma.github.io/logo-website/

Figures & tables

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).
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.
Figure 4: a) : illustration of voxel reward aggregation. b) : visualization of local reward projected to video frames.
Figure 5: Camera trajectories of TrajectoryBench vs. WorldScore.
 Model3D consistencyCameraVideo quality
 PSNR-V ↑\uparrowPSNR-D ↑\uparrowPSNR-GS ↑\uparrowMVCS ↑\uparrowEpipolar ↓\downarrowRPE ∘ ↓\downarrowRPE trans ↓\downarrowVR-VQ ↑\uparrow
Lingbot2Base16.916.016.90.8421.50713.20.1080.21
VideoGPA17.116.117.00.8471.48613.10.1070.22
World-R117.016.016.90.8391.51513.10.1080.21
Global-only (ours)18.017.117.80.8721.30611.20.1020.31
LoGo (ours)19.018.218.40.8941.1118.90.1030.37
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.
 Model3D consistencyCameraVideo quality
 PSNR-V ↑\uparrowPSNR-D ↑\uparrowPSNR-GS ↑\uparrowMVCS ↑\uparrowEpipolar ↓\downarrowRPE ∘ ↓\downarrowRPE trans ↓\downarrowVR-VQ ↑\uparrow
Lingbot2Base13.313.115.60.7192.16120.520.082-0.422
VideoGPA13.313.215.60.7282.42920.150.081-0.409
World-R113.613.315.70.7172.27120.130.070-0.429
Global-only (ours)14.013.816.30.7722.10722.680.086-0.305
LoGo (ours)16.115.817.30.7951.36518.330.077-0.283
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.
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.
 ModelPSNR-V ↑\uparrowPSNR-D ↑\uparrowEpipolar ↓\downarrow
 EMHEMHEMH
Lingbot2Base22.618.215.320.617.614.50.5240.6812.034
VideoGPA22.718.315.420.717.714.60.5330.6841.997
World-R122.718.115.420.617.414.50.5260.6812.046
LoGo23.820.217.621.619.717.00.4900.6011.439
UniWorldBase23.017.216.721.216.416.10.4500.7601.623
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.
Figure 8: Epipolar error vs. length.
 3D ConsistencyCameraVideo Quality
PSNR-V ↑\uparrowPSNR-D ↑\uparrowPSNR-GS ↑\uparrowMVCS ↑\uparrowEpipolar ↓\downarrowRPE ∘ ↓\downarrowRPE trans ↓\downarrowVR-VQ ↑\uparrow
Base15.314.515.70.8122.03413.20.1080.21
Global-only16.515.716.70.8481.72611.20.1020.31
Window17.016.317.10.8561.6669.50.0940.32
Framewise17.016.416.90.8571.53110.70.1010.24
Patchwise16.916.217.00.8651.6608.40.1000.24
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.
 3D ConsistencyCameraVideo Quality
PSNR-V ↑\uparrowPSNR-D ↑\uparrowPSNR-GS ↑\uparrowMVCS ↑\uparrowEpipolar ↓\downarrowRPE ∘ ↓\downarrowRPE trans ↓\downarrowVR-VQ ↑\uparrowHPSv3 ↑\uparrow
Base16.916.016.90.8421.50713.20.1080.215.09
Local-only19.218.218.70.8931.07811.20.1160.324.47
+ interleave19.218.318.60.8971.0728.90.1070.384.94
+ global (final)19.018.218.40.8941.1118.90.1030.375.18
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.
Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

 Model3D consistencyCameraVideo quality
 PSNR-V ↑\uparrowPSNR-D ↑\uparrowPSNR-GS ↑\uparrowMVCS ↑\uparrowEpipolar ↓\downarrowRPE ∘ ↓\downarrowRPE trans ↓\downarrowVR-VQ ↑\uparrow
 EMHEMHEMHEMHEMH
Lingbot2Base22.618.215.320.617.614.521.018.015.70.9180.8780.8120.5240.6812.03413.20.1080.21
VideoGPA22.7 ±\pm 0.0318.3 ±\pm 0.0815.4 ±\pm 0.0420.7 ±\pm 0.0517.7 ±\pm 0.0814.6 ±\pm 0.0521.1 ±\pm 0.0518.1 ±\pm 0.0715.8 ±\pm 0.040.916 ±\pm 0.0020.880 ±\pm 0.0040.820 ±\pm 0.0050.533 ±\pm 0.0110.684 ±\pm 0.0131.997 ±\pm 0.05213.1 ±\pm 0.140.107 ±\pm 0.0010.22 ±\pm 0.004
World-R122.7 ±\pm 0.0618.1 ±\pm 0.0815.4 ±\pm 0.0420.6 ±\pm 0.0817.4 ±\pm 0.0814.5 ±\pm 0.0521.1 ±\pm 0.0618.0 ±\pm 0.0715.7 ±\pm 0.040.917 ±\pm 0.0020.873 ±\pm 0.0050.809 ±\pm 0.0050.526 ±\pm 0.0120.681 ±\pm 0.0122.046 ±\pm 0.05413.1 ±\pm 0.140.108 ±\pm 0.0010.21 ±\pm 0.004
Global (ours)22.9 ±\pm 0.1419.3 ±\pm 0.1116.5 ±\pm 0.0520.8 ±\pm 0.1318.8 ±\pm 0.1115.7 ±\pm 0.0621.4 ±\pm 0.1118.9 ±\pm 0.0916.7 ±\pm 0.050.916 ±\pm 0.0050.908 ±\pm 0.0060.848 ±\pm 0.0050.521 ±\pm 0.0130.649 ±\pm 0.0151.726 ±\pm 0.05711.2 ±\pm 0.170.102 ±\pm 0.0010.31 ±\pm 0.006
Appendix
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.
 Model3D consistencyCameraVideo quality
 PSNR-V ↑\uparrowPSNR-D ↑\uparrowPSNR-GS ↑\uparrowMVCS ↑\uparrowEpipolar ↓\downarrowRPE ∘ ↓\downarrowRPE trans ↓\downarrowVR-VQ ↑\uparrow
Lingbot2Base13.313.115.60.7192.16120.520.082-0.422
VideoGPA13.3 ±\pm 0.1013.2 ±\pm 0.1115.6 ±\pm 0.090.728 ±\pm 0.0142.429 ±\pm 0.15820.15 ±\pm 1.2150.081 ±\pm 0.005-0.409 ±\pm 0.010
World-R113.6 ±\pm 0.1113.3 ±\pm 0.1015.7 ±\pm 0.110.717 ±\pm 0.0172.271 ±\pm 0.15420.13 ±\pm 1.6810.070 ±\pm 0.004-0.429 ±\pm 0.012
Global (ours)14.0 ±\pm 0.1513.8 ±\pm 0.1416.3 ±\pm 0.140.772 ±\pm 0.0162.107 ±\pm 0.15222.68 ±\pm 1.7450.086 ±\pm 0.006-0.305 ±\pm 0.017
LoGo (ours)16.1 ±\pm 0.2215.8 ±\pm 0.2017.3 ±\pm 0.210.795 ±\pm 0.0211.365 ±\pm 0.19818.33 ±\pm 1.5040.077 ±\pm 0.006-0.283 ±\pm 0.018
Appendix
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.
Reward readout3D consistencyCameraVideo quality
PSNR-V ↑\uparrowPSNR-D ↑\uparrowPSNR-GS ↑\uparrowMVCS ↑\uparrowEpipolar ↓\downarrowRPE ∘ ↓\downarrowRPE trans ↓\downarrowVR-VQ ↑\uparrow
 EMHEMHEMHEMHEMH
Base22.618.215.320.617.614.521.018.015.70.9180.8780.8120.5240.6812.03413.20.1080.21
Global22.9 ±\pm 0.1419.3 ±\pm 0.1116.5 ±\pm 0.0520.8 ±\pm 0.1318.8 ±\pm 0.1115.7 ±\pm 0.0621.4 ±\pm 0.1118.9 ±\pm 0.0916.7 ±\pm 0.050.916 ±\pm 0.0050.908 ±\pm 0.0060.848 ±\pm 0.0050.521 ±\pm 0.0130.649 ±\pm 0.0151.726 ±\pm 0.05711.2 ±\pm 0.170.102 ±\pm 0.0010.31 ±\pm 0.006
Window23.3 ±\pm 0.1719.9 ±\pm 0.1117.0 ±\pm 0.0521.2 ±\pm 0.1519.3 ±\pm 0.1116.3 ±\pm 0.0621.6 ±\pm 0.1219.4 ±\pm 0.0917.1 ±\pm 0.050.918 ±\pm 0.0050.914 ±\pm 0.0060.856 ±\pm 0.0050.539 ±\pm 0.0200.635 ±\pm 0.0141.666 ±\pm 0.0539.5 ±\pm 0.170.094 ±\pm 0.0010.32 ±\pm 0.006
Framewise23.5 ±\pm 0.1719.7 ±\pm 0.1117.0 ±\pm 0.0521.3 ±\pm 0.1619.0 ±\pm 0.1116.4 ±\pm 0.0621.5 ±\pm 0.1119.0 ±\pm 0.0916.9 ±\pm 0.050.929 ±\pm 0.0050.918 ±\pm 0.0060.857 ±\pm 0.0050.531 ±\pm 0.0210.617 ±\pm 0.0131.531 ±\pm 0.05310.7 ±\pm 0.170.101 ±\pm 0.0010.24 ±\pm 0.006
Appendix
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.
 RL algorithm3D consistencyCameraVideo quality
 PSNR-V ↑\uparrowPSNR-D ↑\uparrowPSNR-GS ↑\uparrowMVCS ↑\uparrowEpipolar ↓\downarrowRPE ∘ ↓\downarrowRPE trans ↓\downarrowVR-VQ ↑\uparrow
  EMHEMHEMHEMHEMH
 Base23.017.216.721.216.416.120.617.516.90.9000.8980.8710.4500.7601.62313.80.1320.39
DiffusionNFTOurs-global22.8 ±\pm 0.0517.6 ±\pm 0.0517.6 ±\pm 0.0520.8 ±\pm 0.0616.9 ±\pm 0.0717.1 ±\pm 0.0620.5 ±\pm 0.0717.9 ±\pm 0.0617.6 ±\pm 0.040.899 ±\pm 0.0010.901 ±\pm 0.0050.887 ±\pm 0.0030.447 ±\pm 0.0030.773 ±\pm 0.0171.514 ±\pm 0.03913.3 ±\pm 0.140.126 ±\pm 0.0010.37 ±\pm 0.004
LoGlo23.1 ±\pm 0.0517.9 ±\pm 0.0617.9 ±\pm 0.0621.2 ±\pm 0.0717.2 ±\pm 0.0717.4 ±\pm 0.0620.9 ±\pm 0.0818.4 ±\pm 0.0718.0 ±\pm 0.050.903 ±\pm 0.0020.912 ±\pm 0.0040.895 ±\pm 0.0030.446 ±\pm 0.0030.781 ±\pm 0.0171.436 ±\pm 0.03713.9 ±\pm 0.180.130 ±\pm 0.0010.40 ±\pm 0.004
FlowGRPOOurs-global23.2 ±\pm 0.0317.7 ±\pm 0.0317.5 ±\pm 0.0321.3 ±\pm 0.0616.8 ±\pm 0.0416.8 ±\pm 0.0420.9 ±\pm 0.1018.0 ±\pm 0.0417.5 ±\pm 0.030.901 ±\pm 0.0010.902 ±\pm 0.0020.880 ±\pm 0.0020.456 ±\pm 0.0010.788 ±\pm 0.0141.500 ±\pm 0.03513.8 ±\pm 0.070.132 ±\pm 0.0000.37 ±\pm 0.003
Appendix
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.
 Voxel size3D consistencyCameraVideo quality
 PSNR-V ↑\uparrowPSNR-D ↑\uparrowPSNR-GS ↑\uparrowMVCS ↑\uparrowEpipolar ↓\downarrowRPE ∘ ↓\downarrowRPE trans ↓\downarrowVR-VQ ↑\uparrow
  EMHEMHEMHEMHEMH
 Base22.618.215.320.617.614.521.018.015.70.9180.8780.8120.5240.6812.03413.20.1080.21
0.0524.1 ±\pm 0.1620.7 ±\pm 0.1118.6 ±\pm 0.0621.8 ±\pm 0.1620.1 ±\pm 0.1117.8 ±\pm 0.0621.7 ±\pm 0.1319.7 ±\pm 0.0918.2 ±\pm 0.060.932 ±\pm 0.0050.935 ±\pm 0.0070.880 ±\pm 0.0050.466 ±\pm 0.0100.579 ±\pm 0.0141.229 ±\pm 0.0519.2 ±\pm 0.170.108 ±\pm 0.0010.42 ±\pm 0.006
0.124.3 ±\pm 0.1520.5 ±\pm 0.1118.7 ±\pm 0.0621.8 ±\pm 0.1519.9 ±\pm 0.1117.9 ±\pm 0.0721.8 ±\pm 0.1019.6 ±\pm 0.0918.4 ±\pm 0.060.941 ±\pm 0.0050.939 ±\pm 0.0070.885 ±\pm 0.0060.486 ±\pm 0.0140.577 ±\pm 0.0141.235 ±\pm 0.0519.0 ±\pm 0.190.111 ±\pm 0.0010.42 ±\pm 0.006
0.324.6 ±\pm 0.1620.7 ±\pm 0.1118.4 ±\pm 0.0622.2 ±\pm 0.1620.2 ±\pm 0.1117.6 ±\pm 0.0721.7 ±\pm 0.1119.6 ±\pm 0.0918.0 ±\pm 0.060.947 ±\pm 0.0060.940 ±\pm 0.0070.893 ±\pm 0.0060.467 ±\pm 0.0120.581 ±\pm 0.0131.246 ±\pm 0.05110.5 ±\pm 0.180.117 ±\pm 0.0010.42 ±\pm 0.006
Appendix
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.
 3D consistencyCameraVideo quality
PSNR-V ↑\uparrowPSNR-D ↑\uparrowPSNR-GS ↑\uparrowMVCS ↑\uparrowEpipolar ↓\downarrowRPE ∘ ↓\downarrowRPE trans ↓\downarrowVR-VQ ↑\uparrowHPSv3 ↑\uparrow
Base16.916.016.90.8421.50713.20.1080.215.09
Local reward19.2 ±\pm 0.0518.2 ±\pm 0.0518.7 ±\pm 0.040.893 ±\pm 0.0041.078 ±\pm 0.03411.2 ±\pm 0.120.116 ±\pm 0.0010.32 ±\pm 0.0044.47 ±\pm 0.019
+ interleave19.2 ±\pm 0.0518.3 ±\pm 0.0518.6 ±\pm 0.040.897 ±\pm 0.0041.072 ±\pm 0.0348.9 ±\pm 0.130.107 ±\pm 0.0010.38 ±\pm 0.0044.94 ±\pm 0.013
+ global (final)19.0 ±\pm 0.0518.2 ±\pm 0.0518.4 ±\pm 0.040.894 ±\pm 0.0041.111 ±\pm 0.0338.9 ±\pm 0.130.103 ±\pm 0.0010.37 ±\pm 0.0045.18 ±\pm 0.010
Appendix
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.
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.
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.
More in this topic

Camera-Controlled Video Generation

Camera-controlled video generation creates videos that follow specified viewpoints and camera paths while preserving coherent scenes and motion.

See all 142 papers

Explore similar work

CardsList

  1. GeoFlow: Enforcing Implicit Geometric Consistency in Video Generation

    May 18, 2026Jan Ackermann, Shengqu Cai, Boyang Deng +3Physical Consistency in Video GenerationVideo Diffusion Models

    Stanford University · Google DeepMind

  2. WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory

    Sep 21, 2026Wangbo Yu, Kunhao Liu, Wenbo Hu +83D-Aware Video GenerationCamera-Conditioned Video Generation

    ARC Lab, Tencent IEG · Peking University

  3. Track the Noise, Move the World:3D-Grounded Motion-Consistent Noise for Controllable Video Generation

    Jul 2, 2026Long Vu, Tan Ngo, Animesh Karnewar +5Camera-Controlled Video GenerationVideo Diffusion Models

    Qualcomm AI Research · Trinity College Dublin, Ireland