zju3dv

zju3dv / Scal3R

Public

[CVPR 2026 (Highlight)] Scal3R: Scalable Test-Time Training for Large-Scale 3D Reconstruction

zju3dv.github.ioscal3r 3d-reconstruction depth-estimation large-scale
44
1
100% credibility
Found Apr 10, 2026 at 44 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Scal3R is a research tool that reconstructs 3D scenes, camera poses, depth maps, and point clouds from unordered image collections.

How It Works

1
📸 Discover Scal3R

You hear about a cool tool that turns everyday photos into detailed 3D scenes and camera views.

2
🛠️ Set up easily

Run a simple setup script to prepare everything on your computer in minutes.

3
🧠 Download the brains

Grab the ready-to-use smart models with a quick command.

4
✨ Feed in your photos

Point the tool to a folder of your images and launch it with one command.

5
⏳ Watch it work

Sit back as it processes your pictures, building the 3D world step by step.

🌍 Explore your 3D creation

Enjoy camera positions, depth maps, and colorful point clouds ready to view and use.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 44 to 44 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 Scal3R?

Scal3R turns folders of images into 3D reconstructions by predicting camera poses, depth maps, and point clouds via scalable test-time training on long sequences. Drop images into `--input_dir`, run `python -m scal3r.run`, and get `mat.txt` poses, EasyVolcap YAML cams, optional depths/points in Python/PyTorch. It's a CVPR 2026 highlight tackling memory blowup in large-scale SfM, like scanning rooms or streets.

Why is it gaining traction?

Block-wise processing with overlaps and optional loop closure handles thousands of frames without crashing on consumer GPUs, unlike rigid full-sequence methods. Pretrained checkpoints on Hugging Face mean zero training; just install via bash script and infer. Buzz from CVPR 2026 papers GitHub and Reddit threads on scalable 3D recon draws CV folks eyeing scaler alternatives to COLMAP or NerfStudio.

Who should use this?

CV researchers prototyping SfM on video datasets, robotics engineers fusing images for SLAM, AR devs needing quick scene meshes from phone captures. Skip if you want production polish or sub-100-frame toy demos.

Verdict

Grab it for bleeding-edge large-scale 3D if you're tracking CVPR 2026 accepted papers—44 stars and 1.0% credibility scream early days, but inference rocks while eval code lags. Fork the CVPR GitHub template and benchmark against baselines.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.