Research map / Probabilistic and scientific modeling / Molecules and proteins
SubjectDrive: Scaling Generative Data in Autonomous Driving via Subject Control
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SubjectDrive, a subject-controlled multi-view driving video diffusion model (built on Panacea) that injects diverse external vehicle subjects to make generated training data scale for BEV detection and tracking.
Problem
Generative driving-scene data fails to improve perception models as its quantity scales, because generated foreground objects lack diversity.
Approach
Adds subject control to a layout-conditioned video diffusion pipeline via a Subject Prompt Adapter, a Subject Visual Adapter with gated self-attention, and Augmented Temporal Attention, so external subject images (CompCars) can diversify foreground vehicles in generated data.
Key results
- With 69K generated samples plus real nuScenes data, SubjectDrive reaches 50.2 NDS and 37.2 AMOTA, versus 49.2 NDS and 33.7 AMOTA for Panacea at a similar scale.
- Generated-only training reaches 41.1 NDS (88% of real-only performance) and outperforms Panacea by 5.0 NDS.
- Generation quality reaches FVD 124 and FID 15.98 on the nuScenes validation set, the best among the compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase Panacea synthetic data from 46K to 69K | Panacea with StreamPETR · nuScenes val | NDS: 49.3 -> 49.4 | No clear effect |
| increase SubjectDrive synthetic data from 46K to 69K | SubjectDrive with StreamPETR · nuScenes val | NDS: 49.5 -> 50.2 | Helped |
| use SubjectDrive instead of Panacea for synthetic training data | StreamPETR detection · nuScenes val, synthetic only | NDS: 36.1 -> 41.1 | Helped |
| use SubjectDrive instead of Panacea for added synthetic data | StreamPETR tracking · nuScenes val, real+generated | AMOTA: 33.7 -> 37.2 | Helped |
| add SVA (ablation) | SubjectDrive without subject-control modules · nuScenes val, generated samples | NDS: 32.1 -> 34.3 | Helped |
| add SVA (ablation) | SubjectDrive without subject-control modules · nuScenes val, generated samples | AMOTA: 11.4 -> 15.6 | Helped |
| add SPA to SVA (ablation) | SubjectDrive with SVA · nuScenes val, generated samples | NDS: 34.3 -> 36.1 | Helped |
| add SPA to SVA (ablation) | SubjectDrive with SVA · nuScenes val, generated samples | AMOTA: 15.6 -> 17.4 | Helped |
| add ATA to SVA and SPA (ablation) | SubjectDrive with SVA and SPA · nuScenes val, generated samples | AMOTA: 17.4 -> 18.1 | Helped |
| add ATA to SVA and SPA (ablation) | SubjectDrive with SVA and SPA · nuScenes val, generated samples | RECALL: 27.1 -> 28.1 | Helped |
Surprising findings
- Scaling Panacea's generated data from 46K to 69K samples decreased AMOTA slightly (35.0 to 34.9), indicating existing generators do not scale.
Limitations
- Generated-only performance remains below real-data training (41.1 vs 46.9 NDS), so synthetic data does not fully replace real data.
- Evaluation limited to nuScenes with a single perception model (StreamPETR) and low resolution 256x512 with 8-frame clips.
Experimental setup
Datasets: nuScenes CompCars nuImages
Benchmarks: nuScenes
Metrics: NDS mAP mAOE mAVE AMOTA AMOTP MOTA Recall FID FVD
Baselines: Panacea DriveDreamer WoVoGen MagicDrive BEVGen BEVControl
Builds on: Panacea ControlNet latent diffusion models gated self-attention (GLIGEN)
Closest papers by meaning
- DriveGPT: Scaling Autoregressive Behavior Models for Driving (ICML 2025 · similarity 0.87)
- Generating Multimodal Driving Scenes via Next-Scene Prediction (CVPR 2025 · similarity 0.87)
- Panacea: Panoramic and Controllable Video Generation for Autonomous Driving (CVPR 2024 · similarity 0.86)
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.86)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.86)
- Less Language, More Latents: Annotation-Efficient VLAs for Driving (NeurIPS 2026 · 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.