Source-linked AI summary
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
TL;DR
Long-horizon camera-controlled video generation is difficult to evaluate and improve for 3D consistency, while whole-video rewards can miss localized failures. LoGo combines global and spatially localized rewards; across three base models, it improves consistency on challenging benchmarks, and the paper introduces TrajectoryBench for complex-camera evaluation.
Problem
Whole-video scalar rewards lack fine-grained credit assignment for localized 3D-consistency failures in long camera-controlled generations.
Source: §1 · p. 2Method
LoGo post-trains camera-controlled video models with a blend of global and spatially localized 3D rewards.
Source: §1 · p. 2Results
Across three base models, LoGo significantly improves 3D consistency on TrajectoryBench and DL3DV, with up to +2.7dB PSNR and 37% decrease of epipolar error.
Source: §1 · p. 2, §4.3 · p. 8Takeaways & Limitations
TrajectoryBench provides a testbed for long-horizon video generation with complex camera trajectories, while LoGo’s results highlight 3D credit assignment in post-training.
Source: §1 · p. 2, §6 · p. 10LoGo focuses on static scenes, and generation beyond 400 frames remains challenging and may require new memory design.
Source: §6 · p. 10
Abstract
from arXiv · showhide
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/
1 INTRODUCTION
Camera-controlled video generation needs persistent scene structure and objects, but whole-video rewards can miss localized inconsistencies in long rollouts. LoGo combines spatially localized and global rewards, and the paper introduces TrajectoryBench to evaluate long-horizon generation with complex camera trajectories.
1 INTRODUCTION
- Whole-video scalar rewards can overlook localized 3D failures in long generations, even when the rest of the scene structure is sound.The paper notes these challenges in generations up to 400 frames.Source: §1 · p. 2
- LoGo combines spatially localized and global rewards to target local inconsistencies while preserving video quality and camera following.Its 3D reward pinpoints localized errors that a whole-video score can wash out.Source: §1 · p. 2
- TrajectoryBench addresses the lack of long-horizon 3D-consistency evaluation with complex camera trajectories and revisits from new angles.Existing benchmarks focus on 80-frame generations with simple camera commands.Source: §1 · p. 2
- Across three base video models, LoGo outperforms global-reward baselines and achieves state-of-the-art results on two challenging benchmarks.The paper names Lingbot2, Lyra2, and UniWorld as base models.Source: §1 · p. 2
2 RELATED WORK
Prior work improves 3D consistency through geometry-based rewards and other approaches, but much post-training research remains focused on short generations and simple camera trajectories. Other work explores localized rewards, though the paper identifies comprehensive credit assignment as an open focus.
Post-training for 3D consistency of video models
- Geometry-reward post-training methods use epipolar constraints, depth reprojection, or Gaussian splatting, but often target 80-frame videos and simple trajectories.Related approaches also use similar rewards for camera control.Source: §2 · p. 2
Credit assignment in RL
- Localized reward has been explored in editing, omni-model post-training, and robotic video generation, but the paper says these works do not comprehensively explore credit assignment.The passage cites SpatialFlow-GRPO (Yang et al., 2026), LongCat (Team et al., 2026b), and CreFlow (Ni et al., 2026).Source: §2 · p. 3
Long-horizon video generation post-training
- Existing long-horizon post-training targets camera following or reversible actions, while some approaches train on sampled chunks without video-wide credit assignment.Wang et al. (2026c) and DreamX-World 1.0 (Team et al., 2026a) focus on camera following; Gu et al. (2026) studies reversible actions.Source: §2 · p. 3
3 METHOD
LoGo’s method localizes reward in 3D, blends local and global signals, and interleaves consistency training with aesthetic and camera rewards. The recipe is designed to improve consistency while retaining video quality and camera control.
3.1 REWARD LOCALIZATION IN 3D
- The local reward voxelizes a shared 3D scene and assigns each voxel an error from the average reprojection error of its unprojected pixels.Pixel reprojection error combines RGB and depth discrepancies between projections and generated RGB or predicted depth.Source: §3.1 · pp. 3–4
3.2 LOGO: BLENDING LOCAL AND GLOBAL REWARD
- A local reward alone improves consistency but degrades video quality, so LoGo blends normalized local and global rewards to balance consistency and quality.The blended reward expands the consistency-quality Pareto frontier and improves geometry reward without degrading aesthetics.Source: §3.2 · p. 4
3.3 REWARD INTERLEAVING
- LoGo interleaves consistency-reward steps with aesthetic-reward and camera-reward steps to preserve or improve other metrics while improving consistency.The schedule cycles through k steps of LoGo reward, m aesthetic-reward steps, and n camera-reward steps.Source: §3.3 · pp. 4–5
- LoGo adds 0.5% to total per-step training time, with pixel-voxel correspondence accounting for 9 seconds of 1,920 seconds per step.The paper reports the added computation as 2.1% of total reward-calculation time.Source: §3.3 · p. 5
4 EXPERIMENTS
The experiments introduce TrajectoryBench for long-horizon generation with complex camera trajectories, then evaluate LoGo against post-training baselines across three base models. Results compare consistency, camera following, and video quality on TrajectoryBench and DL3DV, with further analyses of difficulty and qualitative failures.
4.1 BENCHMARKS
- TrajectoryBench tests long-horizon generation with trajectories that move beyond the initial view, revisit areas from new angles, and transition into new spaces.It contains 2,000 examples across easy, medium, and hard levels and indoor, outdoor, and transition space types.Source: §4.1 · p. 5
4.2 TRAINING AND EVALUATION SETTINGS
- The evaluation post-trains Lingbot-world-v2, Lyra-2, and UniWorld, measuring 3D consistency, camera following, and video quality against VideoGPA and World-R1.DL3DV uses its original camera trajectories to evaluate camera-control accuracy.Source: §4.2 · pp. 5–6, §4.1 · p. 5
4.3 RESULTS
- Across TrajectoryBench and DL3DV, LoGo improves consistency across three base models while maintaining or improving camera control and video quality; on DL3DV, epipolar error falls by up to 37%.On TrajectoryBench, LoGo increases PSNR-D by 2.2 dB from the Lingbot2 base model and reduces epipolar error by 26%.Source: Table 2 · p. 6, Table 1 · p. 6, §4.3 · p. 8
- Qualitative comparisons show LoGo resolving scene changes, hallucinated objects, floaters, and appearance inconsistencies that remain or shift under baseline methods.Examples include baseline methods retaining a hallucinated object at the Golden Gate Bridge and changing a hallway into a staircase scene.Source: Figure 6 · p. 7, §4.3 · p. 8
- Compared with the global-only variant, LoGo removes local inconsistencies such as a double railing, floaters, and altered chair orientation in reprojection comparisons.The qualitative comparison accompanies the reported importance of local-reward credit assignment.Source: Figure 7 · p. 7, §4.3 · p. 8
- LoGo’s margin over baselines widens as TrajectoryBench difficulty increases, including a Lingbot2 PSNR-D gain of +1 dB on easy trajectories versus +2.5 dB on hard trajectories.The difficulty breakdown reports a widening margin across all three base models.Source: §4.3 · p. 8, Table 3 · p. 8
5 ABLATIONS AND ANALYSES
The analyses test LoGo across video lengths, local-reward assignment strategies, recipe components, voxel granularity, and post-training algorithms. They find stronger gains at longer horizons, favor voxel-based assignment, and show that reward interleaving with local-global blending balances consistency with camera and quality metrics.
Local reward comparisons
- Voxel-based credit assignment yields stronger consistency gains than framewise, 2D patchwise, or 40-frame-window assignment.The analysis notes that imperfect camera following can make a fixed frame or patch correspond to different regions across rollouts.Source: Table 4 · p. 9, §5 · p. 9
Ablating the LoGo recipe
- LoGo’s local-global recipe yields the best overall result: local-only reward maximizes 3D consistency but lowers camera and aesthetics metrics, while blending restores those metrics above the base model.Reward interleaving without local-global blending improves camera and video quality but still degrades HPSv3.Source: Table 5 · p. 9, §5 · p. 9
Video length
- LoGo’s gain widens with video length, reducing epipolar error by 24% at 80 frames and 62% at 400 frames.The base Lingbot2 model’s epipolar errors increase as video length grows from 80 to 400 frames.Source: §5 · p. 9, Figure 8 · p. 9
Robustness to post-training algorithms
- LoGo’s advantage persists when post-training uses FlowGRPO or DRaFT, in addition to the main DiffusionNFT results.The comparisons are reported in the appendix.Source: §5 · p. 9
Voxel granularity
- Finer voxel granularity up to 0.1× P90 scene depth improves both 3D consistency and camera control across the evaluated voxel sizes.The tested range spans 0.05× to 0.7× P90 scene depth.Source: §5 · p. 9
6 CONCLUSION AND LIMITATIONS
The conclusion presents LoGo’s local-global reward blending for post-training camera-controlled video models and reports that it outperforms prior methods. It also introduces TrajectoryBench and identifies limits beyond 400 frames and in dynamic scenes.
6 CONCLUSION AND LIMITATIONS
- LoGo uses local-global reward blending for camera-controlled video post-training and outperforms prior methods through 3D credit assignment.The paper also releases code, checkpoints, and the benchmark.Source: §6 · p. 10
- TrajectoryBench is proposed as a testbed for evaluating long-context camera-controlled video generation.The paper motivates it by the lack of 3D-consistency evaluation for this setting.Source: §6 · p. 10
- Beyond 400 frames, generation remains challenging and may require new memory design; LoGo currently focuses on static scenes.Extending the approach to dynamic scenes is identified as future work.Source: §6 · p. 10
7 AI USE STATEMENT
The authors report using LLMs to polish language and visualize schematic illustrations, and AI coding tools to accelerate standard implementation; they say generative AI was not used for the core idea or paper content.
7 AI USE STATEMENT
- LLMs were used for language polishing and schematic visualization, while AI coding tools helped implement standard code; generative AI was not used for the core idea or paper content.Source: §7 · p. 10
A APPENDIX
The appendix collects additional explanations, experimental and training details, benchmark-construction material, and further analyses.
A APPENDIX
- The appendix adds localized-reward explanations, experimental tables, training and evaluation details, benchmark-construction details, and ablations and analyses.Source: A APPENDIX · p. 14
A.1 RL REWARD FORMULATION WITH LOGO
LoGo makes reward-based credit assignment non-uniform across a sample by assigning latent spatio-temporal patches separate rewards derived from projected voxels.
A.1 RL REWARD FORMULATION WITH LOGO
- For DiffusionNFT, the canonical implementation uses a scalar reward, motivating LoGo’s non-uniform advantage calculation.Source: §A.1 · p. 14
- LoGo replaces a scalar reward with a different reward for each latent spatio-temporal patch, averaging normalized voxel optimality probabilities projected onto that patch.Source: §A.1 · p. 14
- For each current patch, A denotes the voxels projecting onto it; S_i denotes the pixels projecting into voxel i.The FlowGRPO objective can be decomposed similarly.Source: §A.1 · p. 14
A.2 ADDITIONAL EXPERIMENT RESULTS
The additional results provide full difficulty-level TrajectoryBench evaluations and standard errors. LoGo outperforms baselines across all reported geometry metrics and maintains or improves camera control and video quality.
A.2 ADDITIONAL EXPERIMENT RESULTS
- The full TrajectoryBench results break out metrics by easy, medium, and hard splits and report standard errors.The supplementary results include standard errors to demonstrate the statistical significance of metric improvements.Source: §A.2 · p. 14, Table 6 · p. 14
- Additional tables evaluate localization methods, post-training algorithms, voxel granularities, and LoGo recipe ablations.The localization ablation favors voxel localization, with a growing advantage at greater difficulty.Source: §A.2 · p. 14
- Across all three models, LoGo beats baseline methods on all twelve TrajectoryBench geometry metrics while maintaining or improving camera control and video quality.Its strongest advantage is on the hard split.Source: Table 6 · p. 14
A.3 ADDITIONAL TRAINING DETAILS
The appendix reports LoGo training and evaluation details, expands TrajectoryBench’s construction, and illustrates localized rewards and voxel-granularity results.
A.3 ADDITIONAL TRAINING DETAILS
- Training uses 1,700 DL3DV scenes for Lyra2, Lingbot2, and UniWorld, with model-specific trajectory scaling, reward weights, and caption use.Models train until geometry reward convergence, around 16–36 hours on 64 H100s.Source: §A.3 · pp. 14–16
- On DL3DV, LoGo improves 3D consistency over baseline post-training methods across three base models and four metrics, with up to 37% lower epipolar error.Camera control and video quality are maintained.Source: Table 7 · p. 15
- Voxel localization yields the strongest localization improvement, especially on hard tasks, and LoGo’s advantage persists across three post-training algorithms.Voxel scores normalize rollouts in native 3D space, avoiding shifts in represented regions caused by differences in camera following.Source: Table 8 · p. 15, Table 9 · p. 15
- The LoGo recipe combines reward interleaving and local-global blending to achieve the best overall result in the recipe ablation.Local reward alone gives the best 3D consistency but the worst camera and aesthetics metrics; interleaving improves other metrics but still lowers HPSv3 versus the base model.Source: Table 11 · p. 16
A.4 ADDITIONAL EVALUATION DETAILS
- Depth PSNR uses leave-one-out reprojection from every-20th-frame keyframes, excluding the current frame to avoid self-reprojection dominating sparse long-trajectory point clouds.Source: §A.4 · p. 16
A.5 ADDITIONAL DETAILS ON TRAJECTORYBENCH CONSTRUCTION
- TrajectoryBench has 2,000 tasks across easy, medium, and hard splits, spanning indoor, outdoor, and door-through extension scenes.The easy split uses 80-frame, single-action WorldScore tasks; medium includes 240-frame three-action generations, while hard trajectories reach 240 frames and eight camera motions.Source: §A.5 · p. 16
A.6.1 ADDITIONAL VISUALIZATION OF LOGO’S LOCAL REWARD.
- In the visualization, LoGo’s local reward penalizes the hallucinated, splitting tree region, and post-training removes the tree.Source: §A.6.1 · p. 17
A.6.2 VOXEL GRANULARITY ABLATION.
- Finer voxels improve geometry and camera-control metrics up to 0.1× P90 scene depth, with little further gain at finer granularity.The authors hypothesize that large-voxel penalties can incentivize turning the camera less to avoid penalized regions.Source: §A.6.2 · p. 17, Figure 10 · p. 17