Research map / Video / Video generation
Video Generative Models as Geometry Learner
Haosen Yang, Song Jifei, Zhensong Zhang, Xiatian Zhu, Jiankang Deng
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 image reconstruction alongside geometry (co-generate image and geometry) (ablation) | GeoNeXt · NYUv2 depth | AbsRel: 6.5 -> 5.3 | Improved |
| remove CLIP embedding conditioning (ablation) | GeoNeXt · NYUv2 depth | AbsRel: 5.6 -> 5.3 | Improved |
| jointly train depth and normals instead of separate models (ablation) | GeoNeXt · NYUv2 depth | AbsRel: 5.9 -> 5.3 | Improved |
| jointly train depth and normals instead of normal-only model (ablation) | GeoNeXt · iBims-N normals | mean angular error: 17.3 -> 16.4 | Improved |
| swap depth and normal frame order (ablation) | GeoNeXt · NYUv2/ETH3D/iBims-N | AbsRel / mean angular error: NYUv2 AbsRel 5.7 vs 5.3; ETH3D 5.5 vs 5.6 | No clear effect |
| use video diffusion model (GeoNeXt, 59K samples) instead of image diffusion unified model GeoWizard (208K) | unified geometry estimation · KITTI depth | AbsRel: 14.4 -> 8.2 | Improved |
| use GeoNeXt instead of GeoWizard for normals | unified geometry estimation · Sintel normals | mean angular error: 40.3 -> 33.0 | Improved |
| GeoNeXt vs large-scale discriminative Depth Anything on NYUv2 | GeoNeXt · NYUv2 depth | AbsRel: 4.3 -> 5.3 | Hurt |
| increase from 1 step/no ensemble to 5 steps x 5 ensemble (ablation) | GeoNeXt · ETH3D depth | AbsRel: 5.8 -> 5.6 | Improved |
| GeoNeXt vs GeoWizard on KITTI delta1 | GeoNeXt · KITTI depth | delta1: 82.0 -> 92.6 | Mixed |
Closest papers by meaning
- GeoWizard: Unleashing the Diffusion Priors for 3D Geometry Estimation from a Single Image (ECCV 2024 · similarity 0.90)
- G4Splat: Geometry-Guided Gaussian Splatting with Generative Prior (ICLR 2026 · similarity 0.90)
- GeoVideo: Introducing Geometric Regularization into Video Generation Model (NeurIPS 2025 · similarity 0.89)
- Novel View Synthesis with Pixel-Space Diffusion Models (CVPR 2025 · similarity 0.89)
- LaGeM: A Large Geometry Model for 3D Representation Learning and Diffusion (ICLR 2025 · similarity 0.89)
- Generative Perception of Shape and Material from Differential Motion (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.
