📰 News
- 2026.05 — Released DepthVLM-Bench.
- 2026.05 — Released DepthVLM-4B.
🌟 Model Overview
DepthVLM serves as a unified foundation model for both low-level dense geometry prediction and high-level multimodal understanding, while achieving substantially faster inference compared with existing VLM-based approaches such as DepthLM and Youtu-VL.
By attaching a lightweight depth head to the LLM backbone and adopting a two-stage supervision paradigm, DepthVLM transforms a single VLM into a native dense geometry predictor, while preserving its multimodal capabilities and enhancing its spatial reasoning.
🧠 Key Characteristics
-
Native Dense Metric Depth Estimation in VLMs: Directly predicts geometry within the VLM framework.
-
Unified Multimodal Understanding and Geometry Prediction: Generates full-resolution depth maps alongside language outputs in a single forward pass.
-
Efficient Inference: Achieves higher efficiency compared to per-pixel query or coarse token-level outputs.
-
Versatile Application: Supports both indoor and outdoor metric depth estimation.
-
Improved 3D Spatial Reasoning: Moving toward a truly unified foundation model.
🚀 Main Results
Comparison with VLMs
| Benchmark | Ours-8B | Ours-4B | DepthLM-12B | Youtu-VL-4B |
|---|---|---|---|---|
| Argoverse2 | 0.798 | 0.810 | 0.761 | 0.663 |
| Waymo | 0.865 | 0.879 | 0.588 | 0.473 |
| DDAD | 0.813 | 0.818 | 0.654 | 0.342 |
| NuScenes | 0.831 | 0.821 | 0.736 | 0.698 |
| ETH3D | 0.928 | 0.924 | 0.666 | 0.286 |
| ScanNet++ | 0.901 | 0.861 | 0.756 | 0.522 |
| SUN RGB-D | 0.889 | 0.882 | 0.785 | 0.734 |
| IBims-1 | 0.936 | 0.912 | 0.754 | 0.856 |
| NYUv2 | 0.920 | 0.908 | 0.866 | 0.849 |
Bold and underlined indicate the best and second-best results among the compared models. More details can be found in our paper.
Citation
If you find DepthVLM useful for your research or applications, please consider citing our work using the following BibTeX:
@article{yu2026unlocking,
title={Unlocking Dense Metric Depth Estimation in VLMs},
author={Hanxun Yu and Xuan Qu and Yuxin Wang and Jianke Zhu and Lei Ke},
journal={arXiv preprint arXiv:2605.15876},
year={2026}
}