OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Unik3D

UniK3D is a monocular 3D estimation model with released Python training and inference code, pretrained weights and a linked image demo. From a single RGB image it can predict metric depth, camera rays and 3D points in camera coordinates.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Unik3D — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use the linked demo to try the model on images, or use the Python repository for local inference, saved geometric outputs, camera-parameter inputs or custom training work.

Who it is for

  • Computer-vision researchers exploring monocular 3D estimation or depth prediction
  • Developers building image-to-3D or camera-geometry prototypes
  • Practitioners who need to infer depth or point clouds from single-camera imagery

Documented uses

  • Estimate metric depth from one RGB image.
  • Generate camera-ray predictions and 3D points in camera coordinates.
  • Save predicted maps or a point cloud in PLY format.
  • Condition inference on supplied camera parameters or rays.
  • Adapt the model's forward method for custom training data.
Before you choose

Requirements & limits

The repository links to an installation-free Hugging Face demo and provides code for local use. Local setup instructions list Python 3.10+ and CUDA 11.8+ among suggested requirements, with Linux also listed.

  • The model accepts a single RGB image and can return depth, rays and camera-coordinate 3D points.
  • Ground-truth camera parameters or rays can be provided; the README says the model must be in evaluation mode to use input rays.
  • Local inference supports resolution levels from 0 to below 10; higher values increase detail and decrease speed.
  • The project provides training and inference code and pretrained models, but no specific license is supplied in the project information.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Run local inference on an image

Set up UniK3D and use its documented inference script to save image predictions and a point cloud.

Before you start

  • A clone of the UniK3D repository
  • Python 3.10+
  • Linux and CUDA 11.8+ are listed as suggested requirements; the README notes these are not strict hard requirements.
  • An input image and an evaluation config such as configs/eval/vitl.json

Steps

  1. Create and activate a Python virtual environment, or use the documented conda substitutions.

  2. Install the repository and dependencies with pip install -e . --extra-index-url https://download.pytorch.org/whl/cu121.

  3. Run scripts/infer.py with --input, --output and --config-file; add --save and --save-ply to request saved maps and a PLY point cloud.

  4. If needed, provide camera parameters with --camera-path using a JSON file containing the documented name and params fields.

Expected outcome

The inference script is configured to write output maps and, when requested, a PLY point cloud to the chosen output directory.

Guide limitations

  • The README says the listed system requirements are not hard requirements and notes that differences may occur.
  • The optional camera JSON must match a camera model and parameter format documented by the project.
  • Higher resolution levels increase detail but reduce speed.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
18 October 2026
Original website title
lpiccinelli-eth/UniK3D · Official homepage title · retained in its source language
How we review sourcesSuggest a correction