Research map / Video / Video generation
MagicDrive-V2: High-Resolution Long Video Generation for Autonomous Driving with Adaptive Control
Ruiyuan Gao, Kai Chen, Bo Xiao, Lanqing Hong, Zhenguo Li, Qiang Xu
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 MagicDrive with MagicDrive-V2 (DiT, 3D VAE) | MagicDrive · nuScenes video generation (first 16 frames, validation) | FVD: 218.12 -> 94.84 | Improved |
| replace MagicDrive with MagicDrive-V2 | MagicDrive · nuScenes video generation (first 16 frames, validation) | mAP: 11.86 -> 18.17 | Improved |
| replace MagicDrive image model with MagicDrive-V2 | MagicDrive (Img) · nuScenes image generation (validation) | FID: 16.20 -> 20.91 | Hurt |
| replace MagicDrive image model with MagicDrive-V2 | MagicDrive (Img) · nuScenes image generation (validation) | Vehicle mIoU: 27.01 -> 32.73 | Improved |
| add longer videos to low-resolution training (ablation) | MagicDrive-V2 · Training data mixing ablation (nuScenes) | mAP: 10.17 -> 10.51 | No clear effect |
| add higher-resolution videos to training (ablation) | MagicDrive-V2 · Training data mixing ablation (nuScenes) | mAP: 10.17 -> 14.91 | Improved |
| add mixed frame lengths on top of mixed resolutions (ablation) | MagicDrive-V2 · Training data mixing ablation (nuScenes) | FVD: 96.34 -> 99.66 | Hurt |
| add mixed frame lengths on top of mixed resolutions (ablation) | MagicDrive-V2 · Training data mixing ablation (nuScenes) | mAP: 14.91 -> 15.44 | Improved |
| use spatial-temporal 4x downsampling control encoder instead of global temporal reduction (ablation) | MagicDrive-V2 · over-fitting validation loss | validation loss | Improved |
| add MVDiT cross-view block (ablation) | MagicDrive-V2 · multi-view consistency | visual consistency | Improved |
| use CogVAE instead of SD VAE (2D) | video VAE · VAE reconstruction PSNR, 848x1600 17 frames | PSNR: 33.2846 -> 36.0011 | Improved |
| use CogVAE instead of Open-Sora VAE | video VAE · VAE reconstruction PSNR, 424x800 17 frames | PSNR: 30.2779 -> 33.5852 | Improved |
Later papers that compared with this method
6 later accepted papers report MagicDrive-V2 as a baseline in their results tables; 6 of them report a better number on most of the results they share with it.
- Composing Driving Worlds through Disentangled Control for Adversarial Scenario Generation (ECCV 2026 · better on 6 of 6 shared results)
- ConsisDrive: Identity-Preserving Driving World Models for Video Generation by Instance Mask (ICLR 2026 · better on 3 of 3 shared results)
- GeoFlow: Efficient Driving Video Generation via Geometry-Aligned Priors (ECCV 2026 · better on 2 of 2 shared results)
- RLGF: Reinforcement Learning with Geometric Feedback for Autonomous Driving Video Generation (NeurIPS 2025 · better on 4 of 4 shared results)
- InstaDrive: Instance-Aware Driving World Models for Realistic and Consistent Video Generation (ICCV 2025 · better on 3 of 3 shared results)
- Genesis: Multimodal Driving Scene Generation with Spatio-Temporal and Cross-Modal Consistency (NeurIPS 2025 · better on 1 of 1 shared result)
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
- DriveScape: High-Resolution Driving Video Generation by Multi-View Feature Fusion (CVPR 2025 · similarity 0.89)
- DriveDreamer-2: LLM-Enhanced World Models for Diverse Driving Video Generation (AAAI 2025 · similarity 0.89)
- X-Scene: Large-Scale Driving Scene Generation with High Fidelity and Flexible Controllability (NeurIPS 2025 · similarity 0.87)
- Panacea: Panoramic and Controllable Video Generation for Autonomous Driving (CVPR 2024 · similarity 0.87)
- SimGen: Simulator-conditioned Driving Scene Generation (NeurIPS 2024 · similarity 0.87)
- Glad: A Streaming Scene Generator for Autonomous Driving (ICLR 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.
