Research map / 3D vision / Gaussian splatting
Multi-StyleGS: Stylized Gaussian Splatting with Multiple Styles
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.
Multi-StyleGS stylizes 3D Gaussian Splatting scenes with multiple style images by attaching semantic features to Gaussians, matching style regions via bipartite (Hungarian) matching, and using a DINOv2+VGG local-global feature matching loss for multi-view consistency and memory-efficient training.
Problem
Stylizing a 3D Gaussian Splatting scene with multiple reference styles locally (different styles for different objects) while keeping multi-view consistency and reasonable GPU memory use.
Approach
Augment GS with a learnable semantic feature per Gaussian (trained with SAM-derived labels propagated by a video tracker), partition Gaussians by semantic class, match classes to style images with Hungarian bipartite matching on VGG feature means, and apply a nearest-neighbor feature matching loss on concatenated VGG and DINOv2 features so only one class subset is splatted per pass.
Key results
- Multi-view consistency average 0.96 (ours) vs 1.92 for ref-npr and 1.29 for arf in the main table (lower is better).
- Multi-style loss reduces optimized points per iteration, e.g. truck scene about 362K vs 1.09M with NNFM loss.
- User study average score 7.76 for our multi-style results vs 6.93 for arf and 5.73 for ref-npr.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DINOv2 features to VGG feature matching (ablation) | Multi-StyleGS · TNT multi-view consistency (truck, horse, Family, flower) | consistency score: 1.27 -> 1.06 | Helped |
| use semantic multi-style loss instead of NNFM loss (ablation) | Multi-StyleGS · TNT truck | Gaussian points optimized per iteration: 1,087,646 -> 362,548.6 | Helped |
| use semantic multi-style loss instead of NNFM loss (ablation) | Multi-StyleGS · TNT train | Gaussian points optimized per iteration: 756,747 -> 378,561.5 | Helped |
| show multi-style results in addition to single-style results (ablation) | Multi-StyleGS · TNT user study (truck, horse, Flower) | user preference score: 7.50 -> 7.76 | Helped |
| add KNN smooth regularization to semantic segmentation (ablation) | Multi-StyleGS · 3D scene semantic segmentation | segmentation accuracy | Helped |
| add negative entropy and semantic importance filtering together (ablation) | Multi-StyleGS · 3D scene semantic segmentation | segmentation accuracy | Helped |
| add negative entropy regularization without semantic importance filtering (ablation) | Multi-StyleGS · 3D scene semantic segmentation | segmentation accuracy | No clear effect |
| use multi-style transfer instead of single-style transfer (ablation) | Multi-StyleGS · TNT train | styling time: 23.8m -> 22.4m | Helped |
Limitations
- Requires retraining for each new style; no instant style transfer.
- Relies on SAM and video-tracker-generated semantic labels whose quality is not guaranteed.
- Comparisons with lsnerf limited to LLFF, since its code did not work on TnT.
Experimental setup
Datasets: Tanks and Temples LLFF SAM-generated segmentation labels with video tracker propagation
Benchmarks: Tanks and Temples LLFF
Metrics: SIFID multi-view consistency score (following prior work) GPU memory (points per iteration) styling time user study score
Baselines: snerf arf ref-npr lsnerf
Builds on: 3D Gaussian Splatting ARF nearest-neighbor feature matching Hungarian algorithm SAM
Closest papers by meaning
- ReGS: Reference-based Controllable Scene Stylization with Gaussian Splatting (NeurIPS 2024 · similarity 0.93)
- Morpheus: Text-Driven 3D Gaussian Splat Shape and Color Stylization (CVPR 2025 · similarity 0.92)
- CLIPGaussian: Universal and Multimodal Style Transfer Based on Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- SGSST: Scaling Gaussian Splatting Style Transfer (CVPR 2025 · similarity 0.91)
- 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 0.89)
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.