Video generation models have advanced significantly through the latent diffusion transformers trained with rectified flow techniques. Yet these models still struggle with geometric inconsistencies, unstable motion, and visual artifacts that break the illusion of realistic 3D scenes. 3D-consistent video generation could significantly impact numerous downstream applications in generation and reconstruction tasks. We explore how epipolar geometry constraints improve modern video diffusion models. Despite using massive training data, these models fail to capture fundamental geometric principles. We align diffusion models using pairwise epipolar geometry constraints via preference-based optimization, directly addressing unstable trajectories and geometric artifacts through mathematically principled geometric enforcement. Our approach efficiently enforces geometric principles without requiring end-to-end differentiability. Evaluation demonstrates that classical geometric constraints provide more reliable preference signals than the video-quality and geometry metrics based on foundational models. We construct preferences from static scenes with camera motion, where the epipolar metric is most reliable, while the aligned model generalizes to dynamic scenes. By bridging data-driven learning with classical computer vision, we reduce epipolar error by 31% and improve human-rated consistency from 54% to 72% without compromising visual quality.
Figures & tables
Figure 1: Improving 3D Consistency: First and middle frames from the same generated scenes. As the viewpoint changes, the base model exhibits distortions, and visual artifacts, whereas epipolar geometry optimization preserves the structure of the scene and produces more consistent camera motion.
Figure 2: Epipolar Geometry Optimization. Left: In 3D-consistent videos, corresponding points must lie on epipolar lines; we compute the distance from matched points to their expected epipolar lines in sampled frame pairs. Right: Flow-DPO training with a static penalty aligns the model toward geometrically consistent outputs.
Figure 3: Qualitative Evaluation: Each row shows a temporal sequence generated by the base model or the epipolar-aligned model. The first two rows use Wan-2.1-T2V and the last two Wan-2.1-I2V. Across both, the aligned model exhibits fewer distortions and smoother camera motion. Best seen in the videos on the project page https://epipolar-dpo.github.io/ .
3D Consistency Metrics
3D Scene Reconstruction
Human Eval
Method
Sampson Error ↓
Perspective Realism ↑
PSNR ↑
SSIM ↑
LPIPS ↓
Consistency Rate
Baseline
0.190
0.426
22.32
0.706
0.343
54.1%
Ours
0.131
0.428
23.13
0.729
0.315
71.8%
Table 1: 3D Consistency Evaluation: Epipolar aligned model improves 3D Scene Reconstruction and is preferred by human evaluators.
Figure 4: Human Evaluation and Generalization: The left panel measures human judgments of geometric consistency and motion quality. The right panel evaluates generalization to dynamic scenes.
VBench Motion Metrics
VideoReward
Motion Level
Human Eval
Method
Motion Smoothness ↑
Dynamic Degree ↓
Temporal Flickering ↑
Motion Quality ↑
Mean SSIM ↓
Motion Preference Rate
Baseline
0.981
0.751
0.958
50.0%
0.233
18.5%
Ours
0.984
0.710
0.969
69.5%
0.223
53.2%
Table 2: Motion Quality Evaluation: Epipolar aligned model improves motion stability and is preferred by human evaluators despite dynamics-consistency tradeoffs.
Benchmark
Method
Visual Quality
Motion Quality
Text Alignment
Overall
VideoReward
Supervised Finetuning
37.25%
46.75%
53.0%
44.75%
DPO-Epipolar
65.0%
58.5%
50.5%
58.5%
VBench
Supervised Finetuning
35.8%
53.0%
39.4%
35.2%
DPO-Epipolar
61.3%
55.3%
52.0%
57.9%
MiraData9K
Supervised Finetuning
41.5%
42.0%
51.0%
47.0%
DPO-Epipolar
57.0%
58.0%
52.0%
58.5%
Table 3: Comparison vs. Supervised Finetuning on Multi-View Data. Finetuning on real multi-view static scenes Zhou et al. (2018) ; Ling et al. (2024) propagates dataset bias such as scene type or specific camera motion. As a result the model overfits to dataset specifics rather than learning general principles.
Figure 5: Generalization to Dynamic Scenes: The three examples contain both camera motion and independently moving objects: a moving car, a historical street scene, and a person carrying a surfboard. Despite being trained with preferences from static scenes, the epipolar aligned model produces less distortions and more stable trajectories. Best seen in the videos on the project page https://epipolar-dpo.github.io/ .
VideoReward Metrics
Consistency Metrics
Method
VQ
MQ
TA
Overall
Perspective ↑
Sampson ↓
Dynamics ↓
SFT
66.0%
63.0%
54.0%
64.5%
0.427
0.161
0.225
Flow-RWR Liu et al. (2025)
63.5%
60.5%
57.0%
64.0%
0.434
0.174
0.229
DRO Li et al. (2025b)
65.0%
54.0%
50.5%
64.5%
0.410
0.068
0.195
Epipolar-DPO (Ours)
72.0%
71.0%
55.0%
73.0%
0.428
0.127
0.223
Table 4: Win-rate on the VideoReward benchmark comparing different finetuning strategies with geometric consistency metrics and VLM-based metrics (Visual Quality, Motion Quality and Text Alignment).
Figure 6: Descriptor ablation. Win rate versus the baseline when constructing geometric preferences with different feature descriptors. SIFT with Sampson error provides balanced performance across visual quality, motion quality, text alignment, and overall preference, whereas SEA-RAFT achieves higher visual-quality scores but exhibits reward-hacking behavior.
Preference Metric
VideoReward ↑
MET3R ↓
Sampson ↓
Motion Quality Liu et al. (2025)
61.3%
–
0.179
MET3R
–
0.049
0.176
SIFT + Sampson (ours)
64.3%
0.049
0.131
VGGT comparison on the 5K subset
Preference Metric
VQ ↑
MQ ↑
Sampson ↓
VGGT reprojection
53.7%
56.3%
0.105
Table 5: Metric Comparison: SIFT + Sampson outperform VideoReward and MET3R trained models. On the separate VGGT comparison, it achieves comparable perceptual scores and lower Sampson error than VGGT reprojection.
Penalty Weight λ
Dynamic Degree
Motion Quality
Sampson Error
0
0.627
71.5%
0.127
0.1
0.860
23.5%
0.192
10−5
0.629
70.0%
0.130
10−3
0.710
69.5%
0.131
Table 6: Temporal Penalty Weight: Moderate penalty weights improve dynamic degree while preserving motion quality and geometric consistency. Excessively large weight significantly degrades visual quality.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
VBench Visual Metrics
VideoReward
Human Eval
Method
Background Consistency ↑
Aesthetic Quality ↑
Visual Quality ↑
Visual Preference Rate
Baseline
0.930
0.541
-
15.0%
Ours
0.942
0.551
72.0%
52.8%
Appendix
Table 7: Visual Quality and Aesthetic Fidelity Results (Text-to-Video)
Text-to-Video
Method
Visual Quality
Motion Quality
Text Alignment
Overall
Baseline
13.3%
14.4%
24.2%
8.6%
DPO-Epipolar
18.1%
21.8%
25.0%
13.8%
Appendix
Table 8: Win-rate vs. Wan-2.1-14B Wang et al. (2025a) on the VideoReward Liu et al. (2025) benchmark. The Baseline and Epipolar-Aligned Model contain only 1.3B parameters.
Table 10: Robustness to geometric-estimation noise. Pairwise rankings remain stable under realistic correspondence perturbations. The bootstrap experiment removes 10% of correspondences before fundamental-matrix estimation.
Video Group
Median Sampson Error
Median Inlier Ratio
Real static-scene videos
5.187
78.9%
Generated xw
7.393
65.6%
Generated xl
18.937
31.3%
Appendix
Table 11: Score separation across real and generated videos. Real videos receive the lowest median Sampson error, followed by generated winners and generated losers.
Method
Visual Quality ↑
Motion Quality ↑
Overall ↑
Dynamic Degree ↑
Motion (mean SSIM) ↓
Ours
63.1%
65.8%
59.1%
0.80
0.211
VPO
59.1%
70.6%
82.7%
0.65
0.235
VPO + Ours
67.0%
71.9%
83.6%
0.61
0.234
Appendix
Table 12: Prompt Optimization: VPO optimizes prompts while our method improves geometry alignment. The two approaches are complementary and can be combined to achieve both high visual quality and geometric consistency.
Method
Visual Quality ↑
Motion Quality ↑
Sampson Error ↓
Time (20 Steps) ↓
Reward Guidance
52.63%
52.37%
0.168
65.10
Ours
72.0%
69.6%
0.131
54.2
Appendix
Table 13: Inference-Time Reward Guidance: Runtime is measured over 20 denoising steps, with reward guidance applied every 5 steps.
VCIP & TMCC & DISSec, College of Computer Science, Nankai University · Beijing University of Posts and Telecommunications · LIX, École Polytechnique, IP Paris