Research map / 3D vision / 3D generation
PNeSM: Arbitrary 3D Scene Stylization via Prompt-Based Neural Style Mapping
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 PNeSM, which disentangles a NeRF scene's geometry and appearance via a UV style-pattern mapping and stylizes it with a frozen SANet plus a learned visual prompt, enabling arbitrary 3D scene stylization without per-scene retraining.
Problem
Existing NeRF-based 3D scene stylization methods entangle geometry and appearance and must be re-trained for every new scene, so they cannot stylize arbitrary scenes.
Approach
Map the scene appearance into a shared 2D style pattern space (UV mapping with voxel grids and a cycle loss), stylize that space with a frozen pre-trained 2D style transfer network, and adapt it to scene geometry using a single learnable visual prompt inserted at the bottleneck, which generalizes across scenes when trained on a few scenes.
Key results
- On LLFF, PNeSM achieves short-range warped LPIPS 0.0116 versus 0.0125 for ARF and 0.0229 for StylizedNeRF (lower is better).
- Long-range warped LPIPS is 0.0351 for PNeSM versus 0.0353 for ARF and 0.0627 for StylizedNeRF.
- Qualitatively, the method keeps clear geometry on LLFF and Tanks and Temples while matching style references better than ARF, INS, StyleRF, LSNV and Stylizing-3D-Scene.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a visual prompt to the image stylization network (ablation) | PNeSM · 3D scene stylization | visual quality | Helped |
| Use reconstruction-appearance cubemap instead of noise as the stylization network content input (ablation) | PNeSM · 3D scene stylization | stylization quality | Hurt |
Surprising findings
- Directly stylizing the reconstructed appearance cubemap produces abrupt colors, so a noise image is used as content input instead.
Limitations
- Disentanglement is scene-specific and must be run for each new scene, so only the stylization stage is scene-agnostic.
- Evaluation is limited to a few scenes (Fern, Trex, Truck) and a warped LPIPS consistency metric, with no user study.
- Style loss comparison is deferred to supplementary material and not in the main text.
Experimental setup
Datasets: LLFF Tanks and Temples WikiArt
Benchmarks: LLFF Tanks and Temples
Metrics: warped LPIPS short-range consistency warped LPIPS long-range consistency style loss (supp.)
Baselines: StylizedNeRF ARF INS StyleRF LSNV Stylizing-3D-Scene
Builds on: NeRF NeuTex DVGO SANet VPT
Closest papers by meaning
- Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles (NeurIPS 2025 · similarity 0.91)
- FPRF: Feed-Forward Photorealistic Style Transfer of Large-Scale 3D Neural Radiance Fields (AAAI 2024 · similarity 0.88)
- ReGS: Reference-based Controllable Scene Stylization with Gaussian Splatting (NeurIPS 2024 · similarity 0.87)
- Neural 3D Strokes: Creating Stylized 3D Scenes with Vectorized 3D Strokes (CVPR 2024 · similarity 0.87)
- Multi-StyleGS: Stylized Gaussian Splatting with Multiple Styles (AAAI 2025 · similarity 0.87)
- Morpheus: Text-Driven 3D Gaussian Splat Shape and Color Stylization (CVPR 2025 · 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.