Research map / 3D vision / NeRF and radiance fields
Geometry Transfer for Stylizing Radiance Fields
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.
Stylizes the geometry and appearance of TensoRF radiance fields by using a depth map as a style guide, with deformation fields to keep color aligned with shape, plus RGB-D matching, patch-wise loss and perspective augmentation.
Problem
Existing 3D style transfer for radiance fields transfers only color and texture and leaves scene geometry unchanged, ignoring the geometric aspects of style.
Approach
Use a depth map of the style image as a style guide and optimize a deformation grid (offset per 3D point) so the geometry of a radiance field is stylized while color and density are sampled from the original canonical surface; combine with geometry-aware nearest matching, patch-wise style loss, and depth-scaled style augmentation.
Key results
- On trex and fern scenes, the method achieves the lowest SIFID on all RGB, Gray and Depth metrics (e.g., fern Depth 0.28 vs. 0.36 for ARF).
- User study with 22 participants over 12 stylized scenes gives the method an average rank of 1.55, versus 2.58 for ARF, 2.70 for SNeRF and 3.17 for Ref-NPR.
- Depth-only stylization via a deformation grid keeps colors aligned with shape, whereas directly optimizing density produces background colors on updated foreground parts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Geometry Transfer instead of the strongest baseline | TensoRF stylization · trex | RGB SIFID: 1.54 -> 1.43 | Helped |
| use Geometry Transfer instead of the strongest baseline | TensoRF stylization · trex | Gray SIFID: 0.64 -> 0.58 | Helped |
| use Geometry Transfer instead of the strongest baseline | TensoRF stylization · trex | Depth SIFID: 0.51 -> 0.44 | Helped |
| use Geometry Transfer instead of the strongest baseline | TensoRF stylization · fern | RGB SIFID: 1.11 -> 0.81 | Helped |
| use Geometry Transfer instead of the strongest baseline | TensoRF stylization · fern | Gray SIFID: 0.48 -> 0.37 | Helped |
| use Geometry Transfer instead of the strongest baseline | TensoRF stylization · fern | Depth SIFID: 0.36 -> 0.28 | Helped |
| add geometry features to nearest matching (ablation) | RGB-D stylization · RGB-D stylization | stylization diversity and content preservation | Helped |
| add patch-wise optimization (ablation) | RGB-D stylization · RGB-D stylization | pattern clarity and shape accuracy | Helped |
| add perspective style augmentation (ablation) | RGB-D stylization · RGB-D stylization | depth perception | Helped |
Surprising findings
- Optimizing density alone to stylize shape leads to background colors being assigned to deformed foreground regions, motivating the deformation-field design.
Limitations
- Built on TensoRF, so it cannot handle 360-degree unbounded scenes.
- Stylizing from a single style image is ill-posed, since patterns in 3D do not look identical from different viewpoints.
- Evaluated mainly on forward-facing LLFF scenes and a limited set of ScanNet scenes.
Experimental setup
Datasets: LLFF ScanNet
Benchmarks: LLFF ScanNet
Metrics: SIFID (RGB) SIFID (Gray) SIFID (Depth) average user ranking
Baselines: SNeRF ARF Ref-NPR
Builds on: TensoRF ARF Gatys neural style transfer D-NeRF Panoptic Lifting
Closest papers by meaning
- FPRF: Feed-Forward Photorealistic Style Transfer of Large-Scale 3D Neural Radiance Fields (AAAI 2024 · similarity 0.88)
- S-DyRF: Reference-Based Stylized Radiance Fields for Dynamic Scenes (CVPR 2024 · similarity 0.86)
- ReGS: Reference-based Controllable Scene Stylization with Gaussian Splatting (NeurIPS 2024 · similarity 0.86)
- Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles (NeurIPS 2025 · similarity 0.85)
- Geometry in Style: 3D Stylization via Surface Normal Deformation (CVPR 2025 · similarity 0.85)
- Morpheus: Text-Driven 3D Gaussian Splat Shape and Color Stylization (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.