Research map / 3D vision / NeRF and radiance fields
FPRF: Feed-Forward Photorealistic Style Transfer of Large-Scale 3D Neural 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.
Proposes FPRF, a feed-forward photorealistic style transfer method for large-scale NeRFs that uses a stylizable radiance field (K-planes content and semantic fields distilled from VGG and DINO features) with AdaIN and semantic-matched local AdaIN over a clustered multi-reference style dictionary.
Problem
Photorealistic style transfer of large-scale 3D neural radiance fields usually needs per-scene or per-style optimization and does not scale to large scenes, and prior methods do not support multiple style references.
Approach
Distill VGG and DINO 2D features into a blockwise K-planes radiance field so AdaIN statistics can be applied directly to 3D features, plus a DINO-keyed style dictionary of clustered local style codes with softmax style attention for multi-reference, multi-view consistent stylization without per-style optimization.
Key results
- Single-stage training of the stylizable radiance field takes about 1 hr, versus over 5 hrs for StyleRF and over 10 hrs for UPST-NeRF per-scene optimization (as reported in the text).
- Photorealistic stylization results are obtained with about 10 clusters per reference image, and clustering takes no more than 1 sec per reference image.
- Qualitatively, FPRF preserves multi-view color consistency on Mission Bay where 2D PST methods (CCPL, PhotoWCT^2) produce inconsistent colors, and yields more photorealistic results than LipRF, UPST-NeRF and StyleRF on LLFF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add VGG feature distillation loss (ablation) | FPRF · 3D scene style transfer | stylization quality | Helped |
| Use a pre-trained color decoder instead of an MLP trained from scratch during scene optimization (ablation) | FPRF · Arbitrary style reference images | style generalization | Helped |
| Add guided filtering to semantic features before distillation (ablation) | FPRF · 3D scene style transfer | stylization quality | Helped |
| Use FPRF instead of 2D style transfer methods | San Francisco Mission Bay · Large-scale 3D scenes | multi-view color consistency | Helped |
| Use FPRF instead of LipRF | LLFF · Small-scale 3D scenes | color diversity | Helped |
| Use FPRF instead of StyleRF | LLFF · Small-scale 3D scenes | photorealistic quality | Helped |
Surprising findings
- StyleRF, an artistic feed-forward 3D style transfer method, fails to obtain photorealistic results and loses detailed scene structure on LLFF.
Limitations
- Semantic matching performance is bounded by the capability of the DINO semantic encoder (stated in conclusion).
- Evaluation is mainly qualitative in the main text; no quantitative metrics are reported in the provided text.
Experimental setup
Datasets: San Francisco Mission Bay LLFF
Benchmarks: San Francisco Mission Bay LLFF
Baselines: LipRF UPST-NeRF StyleRF CCPL PhotoWCT^2
Builds on: NeRF AdaIN K-planes Block-NeRF DINO multi-view bootstrapping
Closest papers by meaning
- S-DyRF: Reference-Based Stylized Radiance Fields for Dynamic Scenes (CVPR 2024 · similarity 0.91)
- Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles (NeurIPS 2025 · similarity 0.89)
- ReGS: Reference-based Controllable Scene Stylization with Gaussian Splatting (NeurIPS 2024 · similarity 0.88)
- PNeSM: Arbitrary 3D Scene Stylization via Prompt-Based Neural Style Mapping (AAAI 2024 · similarity 0.88)
- Geometry Transfer for Stylizing Radiance Fields (CVPR 2024 · similarity 0.88)
- 3DToonify: Creating Your High-Fidelity 3D Stylized Avatar Easily from 2D Portrait Images (CVPR 2024 · similarity 0.87)
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.