TencentARC

TencentARC / MotionCrafter

Public

[CVPR 2026] MotionCrafter: Dense Geometry and Motion Reconstruction with a 4D VAE

ruijiezhu94.github.ioMotionCrafter_Page 3d 3d-reconstruction 4d 4d-reconstruction crafter
113
4
100% credibility
Found Feb 11, 2026 at 48 stars 2x -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

MotionCrafter transforms ordinary videos into detailed 3D geometry and motion reconstructions viewable in an interactive display.

How It Works

1
🕵️ Discover MotionCrafter

You stumble upon this fun tool while searching for ways to turn everyday videos into interactive 3D scenes.

2
📥 Grab the essentials

Download the ready-made pieces and example videos to get set up quickly.

3
🎥 Pick your video

Choose any video clip from your phone or computer, like a walk in the park or a dancing pet.

4
✨ Create the 3D magic

Press go and watch as it uncovers hidden 3D shapes and smooth movements frame by frame.

5
👀 Dive into the viewer

Open the interactive playground to spin around, zoom in, and see points flowing with motion.

🎉 Your 3D world lives

You now have a beautiful, moving 3D reconstruction ready to explore or share with friends.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 48 to 113 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is MotionCrafter?

MotionCrafter takes a monocular video and outputs dense geometry point maps plus scene flow for every frame, all in a shared world coordinate system--no post-optimization needed. Built in Python with a 4D VAE and diffusion models, it handles reconstruction via simple CLI inference: feed a video.mp4 to run.py and get NPZ files with point maps, valid masks, and motion vectors. Visualize results instantly using Viser.

Why is it gaining traction?

It offers plug-and-play motion guidance for diffusion models and one-shot motion customization, letting you swap in pretrained Hugging Face weights for quick tests on datasets like ScanNet or Sintel. Developers dig the end-to-end pipeline that skips camera calibration hassles, delivering consistent dense motion from raw footage.

Who should use this?

Computer vision researchers benchmarking 4D reconstruction on dynamic scenes, robotics engineers needing monocular motion estimates for SLAM, or AR/VR devs prototyping dense geometry from phone videos.

Verdict

Grab it for research prototypes--pretrained models and eval scripts work out of the box, despite the 1.0% credibility score and 44 stars signaling early maturity. Docs are solid with arXiv backing, but expect tweaks for production.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.