Research map / 3D vision / NeRF and radiance fields
BerfScene: Bev-conditioned Equivariant Radiance Fields for Infinite 3D Scene Generation
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 BerfScene, a BEV-conditioned equivariant radiance field (EG3D-style U-Net generator with SEL layers, padding and low-pass filters) that synthesizes local 3D scenes and stitches them into large or infinite-scale scenes.
Problem
Generating large-scale or infinite 3D scenes with consistent composition, since existing object-level 3D generators do not handle varying object scales and spatial layouts, and composing local scenes often produces jitter and inconsistency.
Approach
Condition a 3D-aware GAN radiance field on a bird's-eye-view map via spatial encoding layers, and enforce translation equivariance with wide BEV margins, positional Fourier features and low-pass filters before downsampling, so local BEV patches can be rendered and stitched consistently.
Key results
- On CLEVR the method reaches FID 0.96 versus 3.61 for CC3D and 4.67 for EG3D.
- On 3D-Front it reaches FID 36.78 versus 42.88 for CC3D, and on Carla FID 40.7 versus 45.2 for CC3D.
- Ablations show removing low-pass filters lowers CLEVR EQT from 22.02 to 18.19 and removing padding from the BEV map raises FID on 3D-Front from 36.78 to 51.30.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the proposed representation with triplane (ablation) | BerfScene · CLEVR | FID: 18.11 -> 0.96 | Helped |
| replace the proposed representation with triplane (ablation) | BerfScene · Front-3D | FID: 39.17 -> 36.78 | Helped |
| replace the proposed representation with 2D-to-3D extrusion (ablation) | BerfScene · CLEVR | FID: 5.60 -> 0.96 | Helped |
| replace the proposed representation with 2D-to-3D extrusion (ablation) | BerfScene · Front-3D | FID: 50.40 -> 36.78 | Helped |
| add padding around the BEV map (ablation) | BerfScene · CLEVR | EQT: 19.01 -> 22.02 | Helped |
| add padding around the BEV map (ablation) | BerfScene · Front-3D | EQT: 13.32 -> 15.76 | Helped |
| add low-pass filters (ablation) | BerfScene · CLEVR | EQT: 18.19 -> 22.02 | Helped |
| add low-pass filters (ablation) | BerfScene · Front-3D | EQT: 14.45 -> 15.76 | Helped |
| use repeated SEL fusion of the BEV map (ablation) | BerfScene · CLEVR | FID: 6.27 -> 0.96 | Helped |
| use repeated SEL fusion of the BEV map (ablation) | BerfScene · Front-3D | FID: 45.90 -> 36.78 | Helped |
| use BerfScene instead of CC3D | 3D scene generation · CLEVR | FID: 3.61 -> 0.96 | Helped |
| use BerfScene instead of CC3D | 3D scene generation · Front-3D | FID: 42.88 -> 36.78 | Helped |
Surprising findings
- Triplane and extruded-plane backbones give worse FID and EQT than the proposed design, with the extruded plane having a higher CLEVR EQT (20.13) than the triplane but much worse 3D-Front FID.
Limitations
- Limited camera views at inference for large-scale scenes, since the generative radiance field mostly learns from the training distribution.
- Only static scenes are supported; dynamic large-scale synthesis is left open.
- Precise attribute control (e.g. color specified in the BEV map) is not guaranteed due to lack of explicit supervision.
Experimental setup
Datasets: CLEVR 3D-Front Carla
Benchmarks: CLEVR 3D-Front Carla
Metrics: FID EQT (equivariance PSNR)
Baselines: StyleGAN2 GSN EG3D CC3D Triplane representation Extruded plane
Builds on: EG3D CC3D StyleGAN2 Spatial Encoding Layer GSN
Closest papers by meaning
- HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis (NeurIPS 2025 · similarity 0.86)
- Free3D: Consistent Novel View Synthesis without 3D Representation (CVPR 2024 · similarity 0.85)
- Latent Radiance Fields with 3D-aware 2D Representations (ICLR 2025 · similarity 0.85)
- NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods (NeurIPS 2025 · similarity 0.85)
- DNGaussian: Optimizing Sparse-View 3D Gaussian Radiance Fields with Global-Local Depth Normalization (CVPR 2024 · similarity 0.85)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.85)
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.