[ICLR2026 Oral (top 1.2%)] DepthLM

Official implementation of "DepthLM: Metric Depth from Vision Language Models".

We show for the first time that VLMs can achieve comparable accuracy with pure vision models on metric depth estimation, with standard text-based SFT and no architecture chagne, i.e., no dense prediction head or regression/regularization loss is needed. Such simplicity allows DepthLM to train a unified VLM to handle various complex 3D understanding tasks such as speed or time estimation, and metric scale camera pose estimation, which require different architecture or hand-crafted pipelines in pure vision models.

Citation

If you find our code useful for your research, please consider citing:

@article{cai2025depthlm,
    title={DepthLM: Metric Depth from Vision Language Models},
    author={Cai, Zhipeng and Yeh, Ching-Feng and Hu, Xu and Liu, Zhuang and Meyer, Gregory and Lei, Xinjie and Zhao, Changsheng and Li, Shang-Wen and Chandra, Vikas and Shi, Yangyang},
    journal={arXiv preprint arXiv:2509.25413},
    year={2025},
}

Contact

Zhipeng Cai, Meta Inc, homepage: https://zhipengcai.github.io/, email: czptc2h at gmail dot com.

Prerequisites

1. run conda create -n DepthLM python=3.12
2. run pip install -r requirements.txt (the code is tested with transformers 4.51.1 version)

Model	Link
DepthLM (Pixtral 12B)	Download 🤗
DepthLM (3B)	(Coming soon!)
DepthLM (7B)	(Coming soon!)

Data Preparation

• For each training/eval dataset, we curate them into
  • A folder containing the images
  • A jsonl file containing the corresponding camera intrinsics and 3D labels
• We provide example data from the iBims1 dataset at examples/ibims1 for quick code run without the need of data preparation. Other images/datasets can use the same code after finishing the data preparation steps.
• Due to legal reasons, we cannot directly release the curated data. However, we provide the data curation code to enable reproduction.
• Checkout each block in prepare_data.sh for the detailed data preparation steps on each dataset.

Eval

• run bash eval.sh <path_to_your_model>

Training

• Download the base model you want to train from here. Our code currently supports Qwen2.5-VL and Pixtral, please see our paper for the corresponding hyper-parameters.
• run bash train.sh <path_to_your_model> <output_path>

Results

Comparison with VLMs

Comparison with pure vision models

Point cloud visualization

Related project

Our follow up project VLM³ has been released! It extends the findings of DepthLM to diverse 3D vision tasks!

License

DepthLM is FAIR CC-BY-NC licensed, as found in the LICENSE file.