Research map / Video / Video generation
GeoFlow: Efficient Driving Video Generation via Geometry-Aligned Priors
Jiazheng Liu, Hang Li, jiawei zhang, Jiahe Li, Xiaohan Yu, Shengyin Fan, Jin Zheng, Xiao Bai
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 |
|---|---|---|---|
| replace Gaussian source with geometry-aligned prior, 5 steps | OpenDWM · NuScenes val, 5 steps | FVD: 123.5 -> 49.2 | Improved |
| replace Gaussian source with geometry-aligned prior, 8 steps vs baseline 40 steps | OpenDWM · NuScenes val | FVD: 38.8 -> 38.6 | Improved |
| add GeoFlow prior to temporal-VAE base | OpenDWM-tvae · NuScenes val, 5 steps | FVD: 204.8 -> 76.7 | Improved |
| add GeoFlow prior to spatial-VAE base | OpenDWM-vae · NuScenes val, 5 steps | FVD: 121.7 -> 49.2 | Improved |
| add GeoFlow prior to stronger base model | UniMLVG · NuScenes val, 5 steps | FVD: 72.6 -> 44.6 | Improved |
| GeoFlow at full 40 steps vs baseline | OpenDWM · NuScenes val, 40 steps | FID/FVD: FID 6.8 -> 6.9, FVD 38.8 -> 34.0 | Mixed |
| add occlusion mask M_occ (ablation) | GeoFlow noise mask · NuScenes val, 10 steps | FVD: 43.6 -> 40.0 | Improved |
| add depth uncertainty mask M_unc on top of M_occ (ablation) | GeoFlow noise mask · NuScenes val, 10 steps | FVD: 40.0 -> 40.2 | No clear effect |
| add dynamic-object mask M_dyn (ablation) | GeoFlow noise mask · NuScenes val, 10 steps | FVD: 40.2 -> 35.0 | Improved |
| use full spatially-adaptive noise mask instead of global uniform noise (ablation) | GeoFlow · NuScenes val | FVD | Improved |
| inject any noise instead of naive warping without noise (ablation) | GeoFlow · NuScenes val | FVD | Improved |
| swap depth model to MapAnything-v1.1 zero-shot (ablation) | GeoFlow · NuScenes val, 10 steps | FVD: 35.0 -> 33.1 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DriveDreamer-2 — DriveDreamer-2: LLM-Enhanced World Models for Diverse Driving Video Generation (AAAI 2025) · better on 2 of 2 shared results
- MagicDrive-V2 — MagicDrive-V2: High-Resolution Long Video Generation for Autonomous Driving with Adaptive Control (ICCV 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- GeoV2V: Geometry-Grounded Video Diffusion Model for Driving Scene Generation (ECCV 2026 · similarity 0.90)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.87)
- Taming generative video models for zero-shot optical flow extraction (NeurIPS 2025 · similarity 0.87)
- StreamFlow: Streamlined Multi-Frame Optical Flow Estimation for Video Sequences (NeurIPS 2024 · similarity 0.87)
- FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video Synthesis (CVPR 2024 · similarity 0.87)
- GeoVideo: Introducing Geometric Regularization into Video Generation Model (NeurIPS 2025 · similarity 0.86)
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.
