Research map / 3D vision / 3D generation
Splat and Distill: Augmenting Teachers with Feed-Forward 3D Reconstruction For 3D-Aware Distillation
David Shavin, Sagie Benaim
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 semantic blending (ablation) | Splat and Distill · ScanNet++ depth (ViT-S) | RMSE: 0.3435 -> 0.3299 | Improved |
| add semantic blending (ablation) | Splat and Distill · ScanNet++ segmentation (ViT-S) | mIoU: 30.99 -> 31.78 | Improved |
| add mask-aware upscaling instead of bilinear (ablation) | Splat and Distill · ScanNet++ depth (ViT-S) | RMSE: 0.3309 -> 0.3299 | Improved |
| use DINO distillation loss instead of cosine loss (ablation) | Splat and Distill · ScanNet++ depth (ViT-S) | RMSE: 0.3310 -> 0.3299 | Improved |
| use EMA-updated teacher instead of frozen teacher (ablation) | Splat and Distill · ScanNet++ depth (ViT-S) | RMSE: 0.3444 -> 0.3299 | Improved |
| render to novel target view instead of context views (ablation) | Splat and Distill · ScanNet++ depth (ViT-S) | RMSE: 0.3332 -> 0.3299 | Improved |
| render to novel target view instead of context views (ablation) | Splat and Distill · ScanNet++ segmentation (ViT-S) | mIoU: 32.08 -> 31.78 | Hurt |
| use manual masks instead of SAM masks (ablation) | Splat and Distill · ScanNet++ depth (ViT-S) | RMSE: 0.3328 -> 0.3299 | Improved |
| use student-teacher distillation instead of direct feature rendering loss (ablation) | Splat and Distill · ScanNet++ depth (ViT-S) | RMSE: 0.3430 -> 0.3299 | Improved |
| add EMA teacher, mask-aware upscaling and blending to basic variant (ablation) | Splat and Distill · ScanNet++ depth (ViT-S) | RMSE: 0.3520 -> 0.3299 | Improved |
| use Splat and Distill instead of Fit3D (ViT-S) | Fit3D · ScanNet++ depth (ViT-S) | RMSE: 0.3506 -> 0.3299 | Improved |
| use Splat and Distill instead of Fit3D (ViT-B) | Fit3D · ScanNet++ segmentation (ViT-B) | mIoU: 34.85 -> 34.07 | Hurt |
Closest papers by meaning
- DUNE: Distilling a Universal Encoder from Heterogeneous 2D and 3D Teachers (CVPR 2025 · similarity 0.90)
- DistillNeRF: Perceiving 3D Scenes from Single-Glance Images by Distilling Neural Fields and Foundation Model Features (NeurIPS 2024 · similarity 0.89)
- Feat2GS: Probing Visual Foundation Models with Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Diffusion 3D Features (Diff3F): Decorating Untextured Shapes with Distilled Semantic Features (CVPR 2024 · similarity 0.87)
- Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (CVPR 2024 · similarity 0.87)
- Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation (NeurIPS 2025 · 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.
