quan-meng

quan-meng / seen2scene

Public

Seen2Scene takes an incomplete real-world 3D scan and generates a complete, coherent 3D scene using visibility-guided flow matching — trained directly on real-world data.

quan-meng.github.ioprojectsseen2scene scene-completion scene-generation
24
0
100% credibility
Found Mar 31, 2026 at 24 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
AI Summary

Seen2Scene is a research method that completes partial real-world 3D scans into full, realistic scenes by learning directly from incomplete data.

How It Works

1
🔍 Discover Seen2Scene

You come across this exciting project on GitHub while looking for ways to fill in missing parts of real-world 3D scans.

2
🎥 Watch the Demo Video

You play the YouTube video and see partial 3D rooms magically turn into full, realistic scenes before your eyes.

3
✨ Explore the Teaser Image

The big preview picture shows off stunning before-and-after examples of cluttered real rooms completed perfectly.

4
📖 Read the Overview

You skim the short summary to understand how it learns from everyday incomplete 3D scans to make them whole.

5
🌐 Visit the Project Page

You click over to the full website to see more examples, details, and possibly try it out yourself.

✅ Master Realistic 3D Completion

Now you know how to create coherent, lifelike 3D environments from partial views, ready for your own creative projects.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 24 to 24 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 seen2scene?

Seen2scene takes an incomplete real-world 3D scan and generates a complete, coherent 3D scene using visibility-guided flow matching, trained directly on real-world data. It solves the problem of turning partial, messy scans from cluttered environments into realistic, fully formed 3D models without relying on synthetic datasets. Developers get access to a research demo via the project page, arXiv paper, and YouTube video, with scenes represented as truncated signed distance fields for practical 3D applications.

Why is it gaining traction?

Unlike alternatives that train on complete synthetic data, seen2scene learns directly from incomplete real-world scans, producing more coherent and realistic completions in complex settings. The visibility-guided flow matching stands out by explicitly handling unknown regions, delivering higher accuracy in generation quality over baselines. Early adopters hook into its flexibility for inputs like 3D layout boxes, text, or partial scans.

Who should use this?

3D reconstruction engineers building AR/VR apps needing quick scene infilling from LiDAR scans. Robotics devs mapping dynamic indoor spaces with sparse sensor data. Computer vision researchers prototyping flow-based generative models for real-world geometry tasks.

Verdict

Skip for production—1.0% credibility score, 24 stars, and just a README with external links signal an early research project lacking code, tests, or docs. Check the project page for demos if you're experimenting with visibility-guided scene completion.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.