Research map / 3D vision / 3D generation
DiffStyle3D: Consistent 3D Gaussian Stylization via Attention Optimization
Yitong Yang, Yinglin Wang, Xuexin Liu, Jing Wang, Hao Dou, Changshuo Wang, Shuting He
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Geometry-Guided Attention across views (ablation) | DiffStyle3D · scene-level 3DGS stylization | CLIP-CONS: 0.121 -> 0.128 | Improved |
| add Geometry-Guided Attention across views (ablation) | DiffStyle3D · scene-level 3DGS stylization | long-range LPIPS: 0.151 -> 0.141 | Improved |
| add geometry-aware mask MG (ablation) | DiffStyle3D · scene-level 3DGS stylization | CLIP-CONS: 0.124 -> 0.128 | No clear effect |
| add centering and normalization to attention losses (ablation) | Attention-Aware Loss · qualitative style transfer convergence | style transfer speed | Improved |
| inject style K and V into rendered Q instead of using style image attention output directly (ablation) | style loss · qualitative | content leakage | Improved |
| use fixed timestep t=1 instead of large, random or decreasing timesteps (ablation) | DiffStyle3D · qualitative | stylization detail | Improved |
| optimize in diffusion latent space with attention loss instead of denoising directions | 3DGS stylization · scene-level 3DGS | FID: 228.6 -> 204.8 | Improved |
| use diffusion attention features rather than CLIP features | 3DGS stylization · scene-level 3DGS | CLIP-S: 0.79 -> 0.65 | Hurt |
| use diffusion attention instead of CLIP features | 3DGS stylization · object-level 3DGS | CLIP-CONS: 0.504 -> 0.412 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- CLIPGaussian — CLIPGaussian: Universal and Multimodal Style Transfer Based on Gaussian Splatting (NeurIPS 2025) · better on 1 of 4 shared results
- SGSST — SGSST: Scaling Gaussian Splatting Style Transfer (CVPR 2025) · better on 2 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles (NeurIPS 2025 · similarity 0.91)
- Morpheus: Text-Driven 3D Gaussian Splat Shape and Color Stylization (CVPR 2025 · similarity 0.90)
- Multi-StyleGS: Stylized Gaussian Splatting with Multiple Styles (AAAI 2025 · similarity 0.90)
- ReGS: Reference-based Controllable Scene Stylization with Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- CLIPGaussian: Universal and Multimodal Style Transfer Based on Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- SGSST: Scaling Gaussian Splatting Style Transfer (CVPR 2025 · similarity 0.88)
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.
