Research map / Image generation / Text-to-image diffusion
FrozenDrive: Zero-Shot Text-Guided Driving Scene Generation and Data Augmentation with Parameter-Free Frozen Diffusion Model
Yuhwan Jeong, HYEONSEONG KIM, Daehyun We, Seonkyu Song, Jinnyeong Yang, Hyun-Kurl Jang, Youngho Yoon, KUK-JIN YOON
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 object-presence ratio loss (upweight rare classes) (ablation) | FrozenDrive · nuScenes val 3DOD ablation (UniAD) | 3DOD mAP: 16.1 -> 21.9 | Improved |
| add object-presence ratio loss (ablation) | FrozenDrive · nuScenes val 3DOD ablation (UniAD) | Motorcycle AP: 0.4 -> 10.0 | Improved |
| add temporal reference self-attention to MISA (ablation) | FrozenDrive · nuScenes val toy ablation (BEV seg, FVD) | FVD: 174.2 -> 144.1 | Improved |
| add multi-view inflated self-attention to TRSA (ablation) | FrozenDrive · nuScenes val toy ablation (BEV seg, FVD) | BEV seg Drivable mIoU: 51.9 -> 56.5 | Improved |
| augment normal-weather training data with FrozenDrive night images | SparseDrive · nuScenes night augmentation (SparseDrive) | 3DOD mAP: 6.62 -> 18.15 | Improved |
| augment with FrozenDrive rain images | SparseDrive · nuScenes rain augmentation (SparseDrive) | planning L2 avg (m): 0.75 -> 0.58 | Improved |
| FrozenDrive rain augmentation versus DriveArena on NDS | SparseDrive · nuScenes rain augmentation (SparseDrive) | NDS: 45.25 -> 44.26 | Hurt |
| use frozen backbone with parameter-free attention instead of learnable cross-attention (SD + CA) (ablation) | FrozenDrive · Lane alignment mIoU | lane mIoU: 33.55 -> 33.58 | No clear effect |
| use frozen backbone for text alignment on unseen snow (ablation) | FrozenDrive · CLIP score, snow | CLIP score: 0.2277 -> 0.2507 (vs DriveArena) | Improved |
| FrozenDrive vs MagicDrive-V2 on temporal consistency | FrozenDrive · VBench temporal consistency | VBench subject consistency: 0.7966 -> 0.7724 | Hurt |
| FrozenDrive vs video-diffusion baselines on FVD | FrozenDrive · nuScenes val generation fidelity (UniAD) | FVD: 81.6 -> 136.8 (MagicDrive-V2) | Hurt |
Closest papers by meaning
- X-Scene: Large-Scale Driving Scene Generation with High Fidelity and Flexible Controllability (NeurIPS 2025 · similarity 0.88)
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.88)
- Generating Multimodal Driving Scenes via Next-Scene Prediction (CVPR 2025 · similarity 0.88)
- Prune2Drive: A Plug-and-Play Framework for Accelerating Vision-Language Models in Autonomous Driving (CVPR 2026 · similarity 0.87)
- SceneCrafter: Controllable Multi-View Driving Scene Editing (CVPR 2025 · similarity 0.87)
- SubjectDrive: Scaling Generative Data in Autonomous Driving via Subject Control (AAAI 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.
