Research map / 3D vision / Gaussian splatting
FruitNinja: 3D Object Interior Texture Generation 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.
Generates interior textures for 3D Gaussian Splatting objects by SDS-based inpainting of canonical cross-sectional views with voxel smoothing and opaque atomic Gaussians, enabling real-time slicing and rendering.
Problem
Interior textures of 3D objects are revealed when objects are sliced or cut, but existing 3D generation, editing and inpainting methods focus on visible surface appearance and cannot produce realistic interior views under arbitrary geometric or topological changes without extra optimization.
Approach
Fill the object interior with Gaussians, progressively inpaint user-specified cross-sectional views with a depth-conditioned diffusion model via SDS, train 3DGS jointly on these slices and surface views, smooth untrained Gaussians via voxel grids, and use opaque, size-capped Gaussians (OpaqueAtom GS) so slicing yields sharp, real-time renderable interiors.
Key results
- FruitNinja achieves CLIP score 33.1 vs 24.6 for PhysGaussian and 32.3 for fine-tuned 2D inpainting on cross-section evaluation.
- FID of 209.2 for FruitNinja vs 520.1 for PhysGaussian and 176.2 for fine-tuned 2D inpainting.
- Cross-view CLIP cosine similarity of 0.96 for FruitNinja vs 0.89 for PhysGaussian and 0.87 for 2D inpainting; rendering is real-time while 2D inpainting needs ~30s per view.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FruitNinja instead of 2D Inpainting (Fine-tuned) | 3D object interior texture generation · Canonical cross-sections | CLIP score: 32.3 -> 33.1 | Helped |
| Use FruitNinja instead of 2D Inpainting (Fine-tuned) | 3D object interior texture generation · Canonical cross-sections | FID: 176.2 -> 209.2 | Hurt |
| Use FruitNinja instead of 2D Inpainting | 3D object interior texture generation · 120 arbitrary cuts | CLIP score: 27.8 -> 29.1 | Helped |
| Use FruitNinja instead of PhysGaussian | 3D object interior texture generation · 120 arbitrary cuts | Cosine similarity: 0.89 -> 0.96 | Helped |
| Add progressive texture refinement (ablation) | FruitNinja · Watermelon and apple cross-sections | Visual fidelity | Helped |
| Add atomic clipping (ablation) | OpaqueAtomGS · 3D object interior textures | Texture realism and convergence | Helped |
| Add uniform and high opacity (ablation) | OpaqueAtomGS · Watermelon slices | Color-transition accuracy | Helped |
| Add voxel smoothing (ablation) | FruitNinja · Cross-sectional views outside the training set | Visual fidelity | Helped |
Surprising findings
- Fine-tuned 2D inpainting obtains a better FID (176.2) than FruitNinja (209.2), which the authors acknowledge as comparable rather than superior.
Limitations
- SDS may introduce high-frequency artifacts that limit photorealism.
- Requires user-specified cross-section angles and 1-4 reference cross-sectional images for fine-tuning.
- Evaluated on only six self-collected objects.
Experimental setup
Datasets: self-collected dataset of 6 objects (watermelon, apple, orange, red velvet cake, loaf bread, pomegranate)
Benchmarks: self-collected cross-section images (FID) CLIP-based cross-section prompts
Metrics: CLIP score FID cosine similarity of CLIP features
Baselines: PhysGaussian 2D Inpainting (Stable Diffusion depth) 2D Inpainting (Fine-tuned)
Builds on: 3D Gaussian Splatting PhysGaussian AtomGS DreamGaussian DreamBooth
Closest papers by meaning
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 2025 · similarity 0.89)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.88)
- DirectTriGS: Triplane-based Gaussian Splatting Field Representation for 3D Generation (CVPR 2025 · similarity 0.87)
- Interactive3D: Create What You Want by Interactive 3D Generation (CVPR 2024 · similarity 0.87)
- DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation (ICLR 2024 · similarity 0.87)
- Spec-Gaussian: Anisotropic View-Dependent Appearance for 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.87)
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.