Research map / Video / Video generation
VIGOR: VIdeo Geometry-Oriented Reward for Temporal Generative Alignment
Tengjiao Yin, Jinglei Shi, Heng Guo, Xi WANG
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 |
|---|---|---|---|
| use pointwise reprojection reward instead of pixel-space reprojection reward for best-of-10 selection (ablation) | CausVid TTS · GB3DV-25k | PSNR / EPI: PSNR 21.07 -> 22.66; EPI 4.549 -> 3.442 | Improved |
| best-of-10 selection with geometry reward vs no selection | CausVid · GB3DV-25k | PSNR: 19.68 -> 22.66 | Improved |
| geometry-reward best-of-10 selection (VIGOR) vs Epipolar reward (ablation) | CausVid · GB3DV-25k | PSNR: 22.45 -> 22.66 | Improved |
| geometry-reward selection does not hurt overall video quality | CausVid · GB3DV-25k VBench Total | VBench Total: 83.33 -> 84.52 | Improved |
| add SFT on reward-selected videos (ablation) | Wan2.1-T2V-1.3B · GB3DV-25k | PSNR: 22.45 -> 23.52 | Improved |
| add SFT on reward-selected videos (ablation) | Wan2.1-T2V-1.3B · GB3DV-25k | dynamic degree (DD): 27.34 -> 15.62 | Hurt |
| add DPO after SFT with Reproj-Pts preference (ablation) | Wan2.1-T2V-1.3B · GB3DV-25k | SSIM: 0.7927 -> 0.7977 | Improved |
| DPO with Reproj-Pts vs Epipolar preference (ablation) | Wan2.1-T2V-1.3B · GB3DV-25k | PSNR / RPT: SSIM 0.7973 -> 0.7977; RPT 1.385 -> 1.424 | Mixed |
| add auxiliary static-penalty loss (lambda) to DPO (ablation) | DPO (Reproj-Pts) · VBench DD | DD delta: +0.78 (lambda 0.1), +1.56 (lambda 10.0) | Improved |
| increase smoothness weight gamma to 10.0 (ablation) | DPO aux loss · VBench MS | MS delta: -0.12 | Hurt |
| replace attention-based geometry-aware sampling with uniform sampling (ablation) | VIGOR reward · VBench DD | DD delta: -4.30 | Hurt |
| use deep-attention layer instead of shallow for sampling (ablation) | VIGOR reward · VBench DD | DD delta: -2.34 | Hurt |
Closest papers by meaning
- VGGRPO: Towards World-Consistent Video Generation with 4D Latent Reward (ECCV 2026 · similarity 0.91)
- VideoGPA: Distilling Geometry Priors for 3D-Consistent Video Generation (ICML 2026 · similarity 0.90)
- Taming Camera-Controlled Video Generation with Verifiable Geometry Reward (ECCV 2026 · similarity 0.90)
- Learning to Correct Geometry in Generated Videos (NeurIPS 2026 · similarity 0.89)
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search (NeurIPS 2025 · similarity 0.87)
- RLGF: Reinforcement Learning with Geometric Feedback for Autonomous Driving Video Generation (NeurIPS 2025 · similarity 0.87)
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.
