---
license: mit
---

# Model Card for Splat and Distill (SnD)

**Splat and Distill (SnD)** is a framework that imparts 3D awareness into 2D Vision Foundation Models (VFMs) by augmenting a teacher network with a feed-forward 3D reconstruction pipeline. It uses 3D Gaussian Splatting (3DGS) to supervise a student model with geometrically consistent features across novel views.



## Model Details

### Model Description

SnD bridges the gap between 2D representation and 3D understanding. It lifts 2D features from a teacher model into a 3D feature field using a feed-forward reconstruction model. These features are then "splatted" onto target views to provide a 3D-consistent supervisory signal for the student.

- **Developed by:** David Shavin, Sagie Benaim
- **Model type:** 3D-Aware Vision Foundation Model (Distillation Framework)
- **Conference:** ICLR 2026
- **License:** MIT
- **Finetuned from model:** DINOv2

### Model Sources

- **Repository:** [https://github.com/davidshavin4/Splat-and-Distill](https://github.com/davidshavin4/Splat-and-Distill)
- **Paper:** [https://arxiv.org/abs/2602.06032](https://arxiv.org/abs/2602.06032)
- **Project Page:** [https://davidshavin4.github.io/Splat-and-Distill/](https://davidshavin4.github.io/Splat-and-Distill/)
- **Blog Post:** [Medium | Splat and Distill](https://medium.com/@davidshavin4/splat-and-distill-augmenting-teachers-with-feed-forward-3d-reconstruction-for-3d-aware-1f2c5e778399)

## Uses

### Direct Use

This model provides 3D-aware semantic features. There are two primary versions available depending on your downstream application:

* **With Blending:** Optimized for **single-view dense estimation tasks**. Use this version for tasks like semantic segmentation, depth estimation, and surface normal estimation.
* **Without Blending:** Optimized for tasks requiring **multi-view correspondence**. Use this version for geometric matching or tasks that rely on consistent feature tracking across different perspectives.

## Bias, Risks, and Limitations

* **Data Bias:** The model was trained using the **ScanNet++** dataset. Consequently, the performance and geometric priors are primarily representative of indoor scene distributions found within that dataset.

## Citation

**BibTeX:**

```bibtex
@misc{shavin2026splatdistillaugmentingteachers,
      title={Splat and Distill: Augmenting Teachers with Feed-Forward 3D Reconstruction For 3D-Aware Distillation}, 
      author={David Shavin and Sagie Benaim},
      year={2026},
      eprint={2602.06032},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={[https://arxiv.org/abs/2602.06032](https://arxiv.org/abs/2602.06032)}, 
}