Research map / 3D vision / 3D generation
PerLDiff: Controllable Street View Synthesis Using Perspective-Layout Diffusion Model
Jinhua Zhang, Hualian Sheng, Sijia Cai, Bing Deng, Qiao Liang, Wen Li, Ying Fu, Jieping Ye, Shuhang Gu
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 PerL-based cross-attention for boxes (ablation) | BEVControl* · NuScenes val | mAP: 16.48 -> 26.07 | Improved |
| Add PerL-based cross-attention for road map (ablation) | BEVControl* · NuScenes val | Road mIoU: 60.74 -> 61.18 | Improved |
| Add both road and box PerL masks (ablation) | BEVControl* · NuScenes val | mAP: 16.48 -> 25.10 | Improved |
| Add both PerL masks (ablation) | BEVControl* · NuScenes val | FID: 13.05 -> 13.36 | No clear effect |
| Add PerLDiff on KITTI trained only on KITTI | BEVControl* · KITTI val | Easy mAP: 0.33 -> 11.04 | Improved |
| Pretrain on NuScenes then fine-tune on KITTI | PerLDiff · KITTI val | Easy mAP: 11.04 -> 13.12 | Improved |
| Increase mask weight lambda from 1.0 to 5.0 (ablation) | PerLDiff · NuScenes val | mAP: 22.30 -> 25.10 | Improved |
| Increase mask weight lambda from 5.0 to 10.0 (ablation) | PerLDiff · NuScenes val | mAP: 25.10 -> 24.98 | No clear effect |
| Replace ConvNext with CLIP road encoder (ablation) | PerLDiff · NuScenes val | NDS: 36.24 -> 36.26 | No clear effect |
| Use GLIGEN-based architecture instead of ControlNet-based (ablation) | PerLDiff · NuScenes val | mAP: 18.07 -> 25.10 | Improved |
| Raise CFG scale from 5.0 to 12.5 (ablation) | PerLDiff · NuScenes val | mAP: 25.10 -> 23.86 | Hurt |
| Add PerL object mask to MagicDrive | MagicDrive · NuScenes val | NDS: 28.79 -> 29.77 | Improved |
Later papers that compared with this method
1 later accepted paper reports PerLDiff as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- DrivePTS: A Progressive Learning Framework with Textual and Structural Enhancement for Driving Scene Generation (CVPR 2026 · better on 5 of 5 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MagicDrive* — MagicDrive: Street View Generation with Diverse 3D Geometry Control (ICLR 2024) · better on 5 of 5 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
- MagicDrive: Street View Generation with Diverse 3D Geometry Control (ICLR 2024 · similarity 0.88)
- StreetCrafter: Street View Synthesis with Controllable Video Diffusion Models (CVPR 2025 · similarity 0.87)
- LSD-3D: Large-Scale 3D Driving Scene Generation with Geometry Grounding (AAAI 2026 · similarity 0.87)
- UrbanGIRAFFE: Representing Urban Scenes as Compositional Generative Neural Feature Fields (ICCV 2023 · similarity 0.86)
- X-Scene: Large-Scale Driving Scene Generation with High Fidelity and Flexible Controllability (NeurIPS 2025 · similarity 0.84)
- DriveGEN: Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation (CVPR 2025 · similarity 0.84)
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.
