Research map / 3D vision / Gaussian splatting
ReGS: Reference-based Controllable Scene Stylization with Gaussian Splatting
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.
Adapts 3D Gaussian Splatting to reference-based scene stylization using a texture-guided densification scheme, depth regularization, and pseudo-view/TCM losses, achieving real-time stylized novel view synthesis.
Problem
Reference-based 3D scene stylization with NeRFs is slow to render, and optimizing the appearance of a pretrained 3D Gaussian Splatting model with fixed geometry fails to capture continuous texture details from the reference image.
Approach
Replace positional-gradient density control with a color-gradient (texture-guided) densification that splits responsible Gaussians into nine tiny Gaussians in a structured layout, while a depth L1 regularizer keeps the original geometry fixed.
Key results
- ReGS reaches Ref-LPIPS 0.202 versus 0.339 for Ref-NPR and 0.394 for ARF.
- ReGS achieves robustness 31.27 PSNR versus 28.11 for Ref-NPR.
- ReGS renders at 91.4 FPS on a single A5000 versus about 16 FPS for NeRF-based baselines (intro claims 134 FPS).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ReGS instead of the strongest baseline, Ref-NPR | Reference-based scene stylization · 12-scene average | Ref-LPIPS: 0.339 -> 0.202 | Helped |
| Use ReGS instead of the strongest baseline, Ref-NPR | Reference-based scene stylization · 12-scene average | Robustness: 28.11 -> 31.27 | Helped |
| Use ReGS instead of the strongest baseline, ARF | Stylized view synthesis · 12-scene average | Speed (fps): 16.5 -> 91.4 | Helped |
| Add texture-guided Gaussian control instead of default positional-gradient density control (ablation) | ReGS · Reference-based scene stylization | Texture detail fidelity: 0.05M newly densified Gaussians -> 0.25M newly densified Gaussians | Helped |
| Add texture-guided Gaussian control rather than optimizing appearance with fixed geometry (ablation) | ReGS · Reference-based scene stylization | Texture detail fidelity | Helped |
| Add depth regularization (ablation) | ReGS · Reference-based scene stylization | Geometry preservation | Helped |
| Add pseudo-view supervision (ablation) | ReGS · Reference-based scene stylization | View consistency | Helped |
Surprising findings
- Positional-gradient density control fails to grow Gaussians at texture-underfitted regions even with 0.25M added Gaussians, while texture-guided control infills most details with 0.05M.
Experimental setup
Datasets: reference-based stylization dataset (12 scenes from Blender, LLFF, Tanks and Temples)
Benchmarks: reference-based stylization dataset (12 scenes)
Metrics: Ref-LPIPS Robustness (PSNR) FPS
Baselines: ARF SNeRF Ref-NPR Naive 3DGS default 3DGS density control
Builds on: 3D Gaussian Splatting Ref-NPR TCM loss Plenoxels
Closest papers by meaning
- Multi-StyleGS: Stylized Gaussian Splatting with Multiple Styles (AAAI 2025 · similarity 0.93)
- Morpheus: Text-Driven 3D Gaussian Splat Shape and Color Stylization (CVPR 2025 · similarity 0.92)
- SGSST: Scaling Gaussian Splatting Style Transfer (CVPR 2025 · similarity 0.90)
- CLIPGaussian: Universal and Multimodal Style Transfer Based on Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 2025 · similarity 0.90)
- Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles (NeurIPS 2025 · similarity 0.90)
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.