nv-tlabs

nv-tlabs / TokenGS

Public

[CVPR'26] TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens

3d-reconstruction novel-view-synthesis
47
1
100% credibility
Found Apr 18, 2026 at 47 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

TokenGS is a research codebase for training neural networks that reconstruct detailed 3D scenes from sparse multi-view images using learnable tokens and Gaussian splatting.

How It Works

1
🔍 Discover TokenGS

You stumble upon TokenGS, a cool tool from NVIDIA researchers that turns a few photos into interactive 3D scenes.

2
💻 Set up your workspace

Download everything to your powerful computer with a good graphics card, and get it ready in minutes.

3
🖼️ Gather your photos

Collect images of a scene from different angles, like everyday objects or spaces, and organize them.

4
🚀 Train your 3D builder

Hit start and watch as it learns to recreate the scene from your photos, building a smart 3D understanding.

5
🔧 Refine for better views

Tweak it with fewer photos to make novel viewpoints sharper and more detailed.

6
👀 Test and preview

Generate new images, depth maps, videos, and 3D files to see how realistic everything looks.

🎉 Explore your 3D world

Enjoy spinning around your recreated scenes, sharing videos or models with friends and colleagues.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 47 to 47 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 TokenGS?

TokenGS is a Python implementation of a CVPR'26 paper that reconstructs 3D scenes as Gaussians from sparse input views. It uses an encoder-decoder to predict learnable tokens, decoupling Gaussian prediction from pixels so the number of primitives stays fixed regardless of image resolution or view count. Users get rendered novel views, depth maps, and PLY exports via simple Accelerate commands like `accelerate launch -m tokengs.train train_dl3dv_base`.

Why is it gaining traction?

Unlike pixel-tied methods, TokenGS fixes token count for consistent scaling across datasets, enabling efficient training on sparse 2-6 view presets. Test-time token tuning boosts eval renders without retraining, and media dumps (PNGs, MP4s, depths) make results easy to inspect. Early adopters praise the self-supervised rendering objective for beating baselines on DL3DV.

Who should use this?

Computer vision researchers benchmarking novel view synthesis on DL3DV. ML engineers prototyping sparse-view 3D Gaussian splatting for AR/VR content generation. Anyone evaluating feed-forward reconstructors needing quick finetunes from base checkpoints.

Verdict

Promising NVIDIA research for token-based 3D recon, but at 47 stars and 1.0% credibility, it's early—docs cover install/train/eval well, yet lacks broad tests or examples. Grab it if you're in Gaussian splatting; otherwise, watch for post-CVPR'26 polish.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.