naver/dust3r

★ 7,132PythonAudience · researcherComplexity · 4/5LicenseSetup · hard

Mindmap

mindmap
  root((DUSt3R))
    What it does
      3D scene reconstruction
      Camera pose estimation
      Point cloud output
    Tech Stack
      Python
      PyTorch
      Hugging Face
      conda
    Use Cases
      Robotics mapping
      Visual effects
      Research experiments
    Input and Output
      Photo sets in
      3D point cloud out
    Setup
      Conda environment
      Model checkpoint
      Docker option

mindmap root((DUSt3R)) What it does 3D scene reconstruction Camera pose estimation Point cloud output Tech Stack Python PyTorch Hugging Face conda Use Cases Robotics mapping Visual effects Research experiments Input and Output Photo sets in 3D point cloud out Setup Conda environment Model checkpoint Docker option

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Build a 3D point cloud of a room or outdoor location by photographing it from multiple angles and running DUSt3R.

USE CASE 2

Use DUSt3R as a camera pose estimation step in a robotics or visual effects pipeline that needs 3D geometry from photos.

USE CASE 3

Experiment with neural 3D reconstruction through a browser-based interactive demo without writing any code.

Tech stack

PythonPyTorchcondaHugging FaceDocker

Getting it running

Difficulty · hard Time to first run · 1h+

Requires conda, a GPU is strongly recommended, and you must download pre-trained model checkpoints from Hugging Face.

Non-commercial use only under CC BY-NC-SA 4.0, you may share and adapt with attribution but cannot use it commercially.

In plain English

DUSt3R is a Python research tool that reconstructs 3D scenes from regular photographs. You give it a set of images taken from different angles of the same place, and it figures out the 3D geometry, estimating where every visible surface sits in three-dimensional space and how the camera was positioned in each shot. This is useful in robotics, visual effects, mapping, and any application where you need a 3D model built from photos alone. The project comes from Naver Labs Europe and was presented at CVPR 2024. The key idea is that the system does not require you to know anything about the camera in advance, such as focal length or lens distortion parameters. It infers all of that from the images. After pairing the input images and running a neural network on each pair, a global alignment step reconciles all the pair-wise estimates into one consistent 3D point cloud. To get started, you clone the repository, set up a Python environment using conda, install the required packages, and download one of the pre-trained model checkpoints from the project's servers or automatically via Hugging Face. Three checkpoint variants are provided, differing in input resolution and decoder architecture. Once the model is ready, you can run an interactive web demo with a single command, select your images through a browser interface, and view the 3D reconstruction when it finishes. A Docker setup is also provided for those who prefer a container-based workflow with optional GPU support. The code can also be used directly in Python scripts. The README includes a short example showing how to load images, create image pairs, run the model, and then call the global aligner to produce the final point cloud. The project is licensed for non-commercial use only under CC BY-NC-SA 4.0.

Copy-paste prompts

Prompt 1

Load a set of 10 photos of my living room into DUSt3R and generate a 3D point cloud I can view and export.

Prompt 2

Integrate DUSt3R's global alignment step into my Python script to merge pair-wise estimates into one consistent 3D reconstruction.

Prompt 3

Run the DUSt3R Docker container with GPU support and reconstruct a scene from a folder of JPEG images.

Prompt 4

Compare the three DUSt3R model checkpoints and choose the best one for my input resolution and available GPU memory.

Open on GitHub → Explain another repo

← naver on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.