Research map / 3D vision / NeRF and radiance fields
S-DyRF: Reference-Based Stylized Radiance Fields for Dynamic Scenes
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.
S-DyRF stylizes dynamic NeRFs (HexPlane-based) from one or a few stylized reference views by generating temporal pseudo-references with a video style transfer method and applying coarse feature-level and fine pseudo-ray color supervision.
Problem
Stylizing dynamic 3D scenes represented as radiance fields from a limited number of stylized reference images, which provide little supervision along the temporal axis, while keeping spatial and temporal consistency.
Approach
Generate temporal pseudo-references from the single stylized reference via a video style transfer model, use them for coarse feature-level style guidance over space and time, and register stylized temporal pseudo-rays (via a temporal reference dictionary over voxels) as explicit fine color supervision.
Key results
- On the Plenoptic Video dataset, Ref-LPIPS with fixed camera viewpoint drops to 0.298 vs 0.374 for Ref-NPR* and 0.353 for Texler et al.
- User study with 56 volunteers prefers S-DyRF over ARF* (88.0%), Ref-NPR* (77.4%), and Texler et al. (70.0%)
- Ablations show removing coarse or fine style transfer degrades stylized novel views/times and style details.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use S-DyRF instead of the strongest baseline | Stylized dynamic radiance fields · Plenoptic Video, fixing camera viewpoint | Ref-LPIPS: 0.353 -> 0.298 | Helped |
| Use S-DyRF instead of the strongest baseline | Stylized dynamic radiance fields · Plenoptic Video, fixing time | Ref-LPIPS: 0.585 -> 0.561 | Helped |
| Use S-DyRF instead of the strongest baseline | Stylized dynamic radiance fields · D-NeRF, fixing camera viewpoint | Ref-LPIPS: 0.593 -> 0.605 | Hurt |
| Use S-DyRF instead of the strongest baseline | Stylized dynamic radiance fields · D-NeRF, fixing time | Ref-LPIPS: 0.621 -> 0.639 | Hurt |
| Remove coarse style transfer (ablation) | S-DyRF · Plenoptic Video | Stylization quality | Hurt |
| Add coarse style transfer (ablation) | S-DyRF · Plenoptic Video | Stylization quality | Helped |
| Remove fine style transfer (ablation) | S-DyRF · Plenoptic Video | Preservation of style details | Hurt |
| Add fine style transfer (ablation) | S-DyRF · Plenoptic Video | Preservation of style details | Helped |
Surprising findings
- Texler et al. video style transfer lags in cross-view/temporal consistency metrics despite strong reference fidelity on some settings
Limitations
- Relies on a pre-trained photorealistic dynamic radiance field and a single stylized reference image or a few references
- Quality depends on the upstream video style transfer model used to generate pseudo-references
- Evaluated on only two datasets
Experimental setup
Datasets: Plenoptic Video dataset D-NeRF dataset
Benchmarks: Plenoptic Video dataset D-NeRF dataset
Metrics: Ref-LPIPS short-range consistency long-range consistency user preference
Baselines: ARF* Ref-NPR* Texler et al.
Builds on: ARF Ref-NPR HexPlane Texler et al. video style transfer
Closest papers by meaning
- FPRF: Feed-Forward Photorealistic Style Transfer of Large-Scale 3D Neural Radiance Fields (AAAI 2024 · similarity 0.91)
- ReGS: Reference-based Controllable Scene Stylization with Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- BLiRF: Bandlimited Radiance Fields for Dynamic Scene Modeling (AAAI 2024 · similarity 0.88)
- Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles (NeurIPS 2025 · similarity 0.88)
- Geometry Transfer for Stylizing Radiance Fields (CVPR 2024 · similarity 0.86)
- PNeSM: Arbitrary 3D Scene Stylization via Prompt-Based Neural Style Mapping (AAAI 2024 · 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.