Research map / Video / Video generation
VGGRPO: Towards World-Consistent Video Generation with 4D Latent Reward
Zhaochong An, Orest Kupyn, Théo Uscidda, Andrea Colaco, Karan Ahuja, Serge Belongie, Mar Gonzalez Franco, Marta Gazulla
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add geometry reprojection reward to motion-smoothness reward (ablation) | VGGRPO · Static+dynamic avg (Wan2.2-5B) | VideoReward visual quality win rate: 55.60 -> 59.57 | Improved |
| add geometry reprojection reward to motion-smoothness reward (ablation) | VGGRPO · Static epipolar error | Sampson epipolar error: 0.104 -> 0.093 | Improved |
| add geometry reprojection reward to motion-smoothness reward (ablation) | VGGRPO · Static+dynamic avg (Wan2.2-5B) | VideoReward motion quality win rate: 63.40 -> 67.21 | Improved |
| use Any4D instead of VGGT as geometry foundation model (ablation) | VGGRPO · Static+dynamic avg (Wan2.2-5B) | VideoReward visual quality win rate: 54.96 -> 59.57 | Improved |
| use Any4D instead of VGGT as geometry foundation model (ablation) | VGGRPO · Static epipolar error | Sampson epipolar error: 0.090 vs 0.093 | No clear effect |
| compute rewards in latent space instead of decoded RGB (ablation) | VGGRPO · Reward computation time | time (s): 54.73 -> 41.33 | Improved |
| compute rewards in latent space instead of decoded RGB (ablation) | VGGRPO · Reward computation memory | peak GPU memory (GB): 76.80 -> 68.57 | Improved |
| add test-time latent reward guidance, training-free | Wan2.2-5B base · Static epipolar error | Sampson epipolar error: 0.142 -> 0.136 | Improved |
| add test-time latent reward guidance, training-free | Wan2.2-5B base · Sampling time | time (s): 44.35 -> 62.60 | Hurt |
| apply VGGRPO post-training | Wan2.2-5B · Standard VBench captions | Imaging quality: 0.6733 -> 0.6861 | Improved |
| apply VGGRPO post-training | Wan2.2-5B · Standard VBench captions | Dynamic degree: 0.4237 -> 0.3962 | Hurt |
| post-train with Epipolar-DPO or VideoGPA instead of nothing | Wan2.2-5B · Standard VBench captions | Imaging quality: 0.6733 -> 0.6353 | Hurt |
Closest papers by meaning
- VIGOR: VIdeo Geometry-Oriented Reward for Temporal Generative Alignment (ECCV 2026 · similarity 0.91)
- VideoGPA: Distilling Geometry Priors for 3D-Consistent Video Generation (ICML 2026 · similarity 0.90)
- Geometry Forcing: Marrying Video Diffusion and 3D Representation for Consistent World Modeling (ICLR 2026 · similarity 0.89)
- G4Splat: Geometry-Guided Gaussian Splatting with Generative Prior (ICLR 2026 · similarity 0.89)
- GeoVideo: Introducing Geometric Regularization into Video Generation Model (NeurIPS 2025 · similarity 0.89)
- PanoWorld: Geometry-Consistent Panoramic Video World Modeling (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
