VLM³: Vision Language Models Are Native 3D Learners

Model	Coming Soon!

Summary

We show that standard VLMs are native 3D learners. We propose VLM³, which without complex data augmentations and any architecture/loss change, can make standard VLMs:

• Surpass SpatialRGPT on object-level 3D understanding (both qualitative and quantitative in SpatialRGPT-bench), without using extra encoders.
• Match UnidepthV2 and Moge-2 on metric depth estimation, improving the accuracy of DepthLM from 0.84 to 0.9;
• Surpass DKM and RoMa for pixel correspondence estimation;
• Match DepthAnything3 and surpass VGGT for camera pose estimation;

VLM³ opens up a new paradigm for simple and scalable 3D learning. Now you dont need to spend a year designing:

• complex models with different backbones, prediction heads, routings.
• complex losses for different prediction heads, balancing weights for different losses
• complex data augmentations like image cropping, rotatin, translation, appearance augmentation etc.

All you need to do is collect data, and scale the training with a standard VLM!

Our findings provide a new perspective on what is and is not necessary for 3D vision:

• Large models, task-specific architectures, losses, data-augmentations, and even the regression formulation that sets the foundation of most SOTA 3D expert vision models, are all not necessary conditions for effective 3D learning.
• A generalist foundation model (VLM) with unified output domain (text) + data scaling are sufficient.

Method Overview

Given the input images, VLM³ first resizes them so that the focal length is the same for all input images (e.g., 1000 pixels). This solves camera ambiguity without the need for adding extra VLM encoders/modules. To refer to an object or pixel, VLM³ simply uses text with the pixel range normalized (e.g., [0, 2000) or [0, 1000)) for both horizontal and vertical axes. This requires no architecture or marker rendering, and makes VLM³ much more flexible and scalable. Standard VLM architectures and text-based training (SFT) are used to train the model.

Results

Contact

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

Quickstart

Install transformers to do inference with our model.

pip install transformers>=5.4.0

Since VLM³ maintain the architecture of the base model (Qwen3-vl-4B), we can call the model for inference the same as the original VLM.

Using 🤗 Transformers to Chat

Here we show a code snippet to show you how to use the chat model with transformers:

from transformers import AutoModelForImageTextToText, AutoProcessor

# default: Load the model on the available device(s)
model = AutoModelForImageTextToText.from_pretrained(
    "facebook/VLM3-depth", dtype="auto", device_map="auto"
)

processor = AutoProcessor.from_pretrained("facebook/VLM3-depth")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://github.com/facebookresearch/VLM3/blob/main/sample_data/depth.jpeg",
            },
            {
                "type": "text",
                "text": (
                    f"Given this image, how far is the point at coordinates ({norm_x}, {norm_y}) "
                    "from the camera? The coordinates are in normalised [0, 2000] format relative "
                    "to image width and height. Output the thinking process in <think> </think> "
                    "and final answer (the meter number only, without the unit) in <answer> </answer> tags."
                ),
            },
        ],
    }
]

# Preparation for inference
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt"
)
inputs = inputs.to(model.device)

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

Please check this Cookbook for detailed examples on how to use different checkpoints for different tasks.

Citation

@article{cai2026vlm3,
    title={VLM³: Vision Language Models Are Native 3D Learners},
    author={Cai, Zhipeng and Liu, Zhuang and Xiong, Yunyang and Liu, Zechun and Vikas, Chandra and Shi, Yangyang},
    journal={arXiv preprint arXiv:xxxx.yyyy},
    year={2026},
}

Related projects

This work is largely motivated by our previous project DepthLM

@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},
}

License

VLM³ is FAIR CC-BY-NC licensed, as found in the LICENSE file.