Research map / 3D vision / 3D generation
Neural 3D Strokes: Creating Stylized 3D Scenes with Vectorized 3D Strokes
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.
Neural 3D Strokes represents a 3D scene as a differentiable field of vectorized 3D strokes (primitives and splines) optimized from multi-view images to produce stylized novel views.
Problem
Stylizing 3D scenes geometrically and aesthetically, since prior NeRF stylization methods only alter color and leave geometry unchanged.
Approach
Replace the MLP-based NeRF field with a stroke field of parameterized 3D strokes whose SDFs are converted to density via a Laplace-CDF smoothing with cone-traced adaptive width, plus an error-field-guided stroke initialization and update scheme that mitigates vanishing gradients.
Key results
- Full method achieves 23.13 PSNR, 0.825 SSIM and 0.171 LPIPS in the ablation on novel view synthesis, outperforming constant-delta and no-error-field variants.
- Overlay composition gives the best reconstruction quality among overlay, max, and softmax compositions, with softmax usable when order invariance is needed.
- Demonstrates geometric stylization with box, ellipsoid, and cubic Bezier strokes, plus color transfer via VGG style loss and text-driven drawing via CLIP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use adaptive δ instead of constant low δ (ablation) | Neural 3D Strokes · Novel view synthesis | PSNR: 21.30 -> 23.13 | Helped |
| use adaptive δ instead of constant high δ (ablation) | Neural 3D Strokes · Novel view synthesis | LPIPS: 0.235 -> 0.171 | Helped |
| add adaptive δ rather than omit it (ablation) | Neural 3D Strokes · Novel view synthesis | PSNR: 22.50 -> 23.13 | Helped |
| add the error field for stroke initialization (ablation) | Neural 3D Strokes · Novel view synthesis | PSNR: 21.79 -> 23.13 | Helped |
| use overlay composition instead of max or softmax composition (ablation) | Neural 3D Strokes · Scene reconstruction | reconstruction quality | Helped |
Surprising findings
- Cubic Bezier and ellipsoid strokes give the best reconstruction fidelity among stroke types, and view dependency contributes little to stroke-based scenes.
Limitations
- Requires manual design of stroke shapes and appearances.
- May need numerous strokes for very complex scenes due to local minima.
- Blurry or abstract results at constant delta; no global optimal-transport loss in 3D yet.
Experimental setup
Datasets: Blender synthetic objects LLFF face-forwarding scenes
Benchmarks: Blender synthetic scenes LLFF
Metrics: PSNR SSIM LPIPS
Baselines: diffvg Stylized Neural Painting
Builds on: NeRF VolSDF Mip-NeRF diffvg
Closest papers by meaning
- Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles (NeurIPS 2025 · similarity 0.87)
- PNeSM: Arbitrary 3D Scene Stylization via Prompt-Based Neural Style Mapping (AAAI 2024 · similarity 0.87)
- 3DToonify: Creating Your High-Fidelity 3D Stylized Avatar Easily from 2D Portrait Images (CVPR 2024 · similarity 0.87)
- Real3D: The Curious Case of Neural Scene Degeneration (AAAI 2024 · similarity 0.86)
- S-DyRF: Reference-Based Stylized Radiance Fields for Dynamic Scenes (CVPR 2024 · similarity 0.86)
- FPRF: Feed-Forward Photorealistic Style Transfer of Large-Scale 3D Neural Radiance Fields (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.