Generated videos often lose geometric consistency under camera motion: structures warp under viewpoint change, parallax becomes inconsistent, and smooth camera movement may collapse into shot-like transitions. We argue that this failure is largely a data problem. In web-scale video corpora, most visible motion comes from dynamic objects captured by static or weakly moving cameras, leaving only sparse supervision for viewpoint-induced appearance changes. To address this, we propose Video Geometry Corrector (VGC), a post-hoc video-to-video model trained on synthetic paired supervision. Starting from real videos with diverse camera motion and pose annotations, we construct VG-Pairs with a synthetic paired-supervision data generation pipeline. Each video clip is treated as a geometry-consistent target and paired with a distorted counterpart synthesized by a pretrained image-to-video inpainting model using trajectory-adaptive keyframes. VGC learns to map the distorted input to the geometry-consistent target without modifying the source generator, and uses optical flow as an auxiliary cue to help disentangle camera motion from object motion. Experiments on pose-conditioned and prompt-driven benchmarks show that VGC improves geometric consistency across multiple source generators, and scaling to a larger backbone further extends the framework toward more general video refinement.
Figures & tables
Figure 1: Video Geometry Corrector ( VGC ) corrects geometry distortions in generated videos. Under large camera motion, modern video generators often produce warped structures, inconsistent parallax, and drifting scene layouts (red boxes). VGC refines these outputs into geometry-consistent videos. When scaled to a larger backbone, it further extends to general video refinement, including dynamic scenes with distorted human or object motion. Project page: https://weijielyu.github.io/VGC .
Figure 2: Effect of adaptive keyframe selection. The red arrows indicate a constant 3D direction while the camera is moving. For clips with large, nonlinear camera motion, conditioning the image-to-video model only on the first and last frames often produces a shortcut trajectory that deviates substantially from the ground-truth path (top; about 45∘ rotation versus about 225∘ in the ground truth). Our adaptive keyframe selection inserts intermediate keyframes, generates shorter segments conditioned on their boundary keyframes, and stitches them together. This produces geometry-distorted videos that remain close to the target camera trajectory and are therefore suitable as paired training data (middle), compared with the ground-truth supervision video (bottom).
Figure 3: Overview of Video Geometry Corrector ( VGC ). Starting from a real video with diverse camera motion, we select trajectory-adaptive keyframes and use them to condition an image-to-video inpainting model, which produces a geometry-distorted video that serves as the training input. We then train VGC to map this distorted input to the geometry-consistent target video with a reconstruction loss. Optical flow is computed from the distorted video and used as an additional conditioning signal for VGC , providing motion cues that help disentangle camera motion from scene dynamics. The 3D reconstruction is shown here only for visualization and is not used in the pipeline.
Method
CogVideoX-5B T2V
CogVideoX 1.5-5B T2V
CogVideoX-5B I2V
PSNR ↑
SSIM ↑
LPIPS ↓
MSE ↓
3DCS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
MSE ↓
3DCS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
MSE ↓
3DCS ↓
Baseline
15.6668
0.5389
0.5285
0.0333
0.8617
15.9062
0.4644
0.5449
0.0289
0.8342
17.3189
0.5543
0.5760
0.0218
0.7944
Baseline + VGC
16.6940 (+1.0272)
0.5519 (+0.0130)
0.5733 (+0.0448)
0.0264 (-0.0069)
0.8376 (-0.0241)
17.7342 (+1.8280)
0.5284 (+0.0640)
0.5936 (+0.0487)
0.0192 (-0.0097)
0.7859 (-0.0483)
17.9989 (+0.6800)
0.5456 (-0.0087)
0.6284 (+0.0524)
0.0184 (-0.0034)
0.8121 (+0.0177)
VideoGPA Du et al. (2026)
14.6318
0.5210
0.5065
0.0428
0.9346
14.4002
0.4175
0.5531
0.0408
0.9614
15.7203
0.5240
0.5749
0.0320
0.8952
VideoGPA Du et al. (2026) + VGC
15.9092 (+1.2774)
0.5413 (+0.0203)
0.5546 (+0.0481)
0.0319 (-0.0109)
0.8737 (-0.0609)
16.8430 (+2.4428)
0.5039 (+0.0864)
0.5907 (+0.0376)
0.0235 (-0.0173)
0.8258 (-0.1356)
17.3053 (+1.5850)
0.5379 (+0.0139)
0.6216 (+0.0467)
0.0225 (-0.0095)
0.8463 (-0.0489)
Table 2: Quantitative results on DL3DV Ling et al. (2024) . We evaluate VGC on three CogVideoX Yang et al. (2025) source generators in both text-to-video and image-to-video settings. VGC consistently improves PSNR, SSIM, and MSE over the corresponding baselines, and further improves VideoGPA Du et al. (2026) when applied as a post-hoc corrector. These results show that VGC transfers across different generation settings and can be flexibly combined with other geometry-aware refinement methods.
Figure 5: Visual results on DL3DV Ling et al. (2024) . The first two rows show image-to-video results. In the baseline CogVideoX Yang et al. (2025) outputs, the streetlight is duplicated, while VideoGPA Du et al. (2026) produces a distorted and tilted streetlight (red boxes). VGC corrects these geometric errors and restores a more plausible structure. The last row shows a text-to-video example, where VGC better preserves the original appearance of the sculpture, while VideoGPA changes its structure and other scene details.
Figure 6: Comparison of model sizes. In static scenes (top), both VGC 1.3B and 14B correct geometric distortions produced by the baseline model that SeedVR2 Wang et al. (2025b) and VEnhancer He et al. (2024) fail to restore (red boxes). The 14B model further improves the geometry of moving humans and objects in dynamic scenes (bottom).
Method
Subject Consistency
Background Consistency
Temporal Flickering
Motion Smoothness
Dynamic Degree
Aesthetic Quality
Imaging Quality
DOVER Wu et al. (2023)
CogVideoX-I2V Yang et al. (2025)
Baseline
0.9480
0.9455
0.9742
0.9918
0.30
0.5095
69.6945
9.54
SeedVR2 Wang et al. (2025b)
0.9447
0.9476
0.9603
0.9829
0.30
0.5313
71.1767
8.31
VEnhancer v2 He et al. (2024)
0.9506
0.9602
0.9749
0.9920
0.31
0.5237
69.3742
10.31
VGC (Ours)
0.9483
0.9450
0.9772
0.9891
0.66
0.5011
72.7654
10.50
HunyuanVideo 1.5-I2V Team (2025)
Table 3: General video refinement results on SpatialVID Wang et al. (2025a) . We compare VGC with general video refinement and enhancement methods Wang et al. (2025b) ; He et al. (2024) on dynamic videos generated by CogVideoX Yang et al. (2025) , HunyuanVideo Team (2025) , and Wan Wan et al. (2025) . Although VGC is designed for geometry correction under camera motion rather than generic enhancement, it achieves competitive performance across multiple perceptual and temporal metrics, and is particularly strong in dynamic degree, imaging quality, and DOVER Wu et al. (2023) . These results suggest that the proposed correction framework extends beyond static-scene geometry and remains effective in more general settings.
Figure 7: General video refinement results. VGC corrects several geometric failures produced by the baseline models. In scene 1, HunyuanVideo Team (2025) changes the cabinet into a glass door, while VGC restores the correct scene structure and camera motion, placing the door on the right side of the cabinet. In scene 2, HunyuanVideo generates a camera trajectory that passes through the wrong side of the doorway, whereas VGC recovers a plausible path. In scenes 3 and 4, VGC restores a consistent appearance and recovers fine geometric structure, such as the wooden supports. In the dynamic scenes, VGC better preserves identity and motion continuity: in scene 5, it maintains the same man in the leather jacket across frames and recovers a plausible motion trajectory, whereas the Wan baseline fails to infer coherent human motion; in scene 6, it restores the missing cyclist produced by CogVideoX Yang et al. (2025) . The green boxes further show that VGC enhances local detail while remaining more faithful to the source video than SeedVR2 Wang et al. (2025b) and VEnhancer He et al. (2024) . Please zoom in for details.
Generating geometrically consistent videos remains an open challenge: text-to-video diffusion models trained on web-scale data treat geometry only implicitly, leading to object deformation, texture drift, and non-rigid backgrounds under camera motion. Existing solutions either improve consistency as a byproduct, apply only to static scenes or realign the latent space of the model completely. We introduce a geometry-consistency reward that directly measures whether motion in a generated video is compatible with a coherent scene. Our key insight is that in physically consistent videos, background motion should be explainable by rigid camera-induced flow, while independently moving objects should preserve appearance identity along motion trajectories. We operationalize this using optical flow, depth--pose predictions, and feature-based correspondence to separate rigid and dynamic regions and evaluate their respective consistency. Integrating this reward with reinforcement fine-tuning transforms geometric consistency from an emergent property into an explicit optimization objective for video generators. The approach is model agnostic and applies to diverse dynamic scenes containing both camera and object motion. Experiments show substantial reductions in temporal geometric artifacts over strong baselines while preserving perceptual quality. Code and model weights are published.
We introduce GeCo, a geometry-grounded metric for jointly detecting geometric deformation and occlusion-inconsistency artifacts in static scenes. By fusing residual motion and depth priors, GeCo produces interpretable, dense consistency maps that reveal these artifacts. We use GeCo to systematically benchmark recent video generation models, uncovering common failure modes, and further employ it as a training-free guidance loss to reduce deformation artifacts during video generation.
Large-scale video diffusion models often fail to preserve 3D structure over time, causing geometric drift and implausible motion under viewpoint changes. Existing methods usually enforce geometric consistency by using explicit geometry reconstructions, such as depth maps, point clouds, or reconstructed 3D structures, to define conditions, supervision, or reward signals, making the generator sensitive to errors from upstream geometry pipelines. We propose VideoWeave, a latent-space post-training framework that uses implicit geometry-model features to constrain the generative distribution, providing a more flexible and non-rigid form of guidance that mitigates the impact of reconstruction errors from geometry models. Specifically, VideoWeave adapts these features into geometry latents and jointly models them with video latents in a shared denoising space, allowing geometry to shape the generative distribution during training. To support this process, we build GeoVid-80K, an 80K-video dataset with paired appearance and geometry representations. Experiments on text-to-video and image-to-video generation show that VideoWeave improves geometric coherence while preserving strong visual quality. VideoWeave project page at https://videoweave.github.io/
Xunzhi Xiang, Zixuan Duan, Yabo Chen +8
1Nanjing University · Institute of Artificial Intelligence, China Telecom (TeleAI) · 3Chinese University of Hong Kong, Shenzhen