mohammadasim98

mohammadasim98 / scenetok

Public

[CVPR '26] SceneTok: A Compressed, Diffusable Token Space for 3D Scenes

geometric-rl.mpi-inf.mpg.descenetok 3d-reconstruction diffusion-models novel-view-synthesis scene-autoencoding scene-generation
62
1
100% credibility
Found Feb 26, 2026 at 48 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
AI Summary

SceneTok is a research project that transforms sets of images from 3D scenes into a compact form usable for generating new viewpoints or entire scenes, with tools coming soon.

How It Works

1
🔍 Discover SceneTok

You come across this cool project while reading about new ways to work with 3D scenes online or at a conference.

2
🌐 Check out the page

You visit the project's website or GitHub to see what it's all about.

3
💡 Grasp the big idea

You learn how it squeezes photos from different angles of a room or object into a tiny, magical summary that can dream up new views or even whole new scenes.

4
📖 Explore the details

You read the simple overview and peek at the pretty diagrams showing it in action.

5
⏳ Stay tuned

You note the upcoming releases for different scene collections and get excited for what's next.

6
📎 Save for later

You grab the citation info to reference this neat work in your own projects or notes.

🎉 All set to play

Soon you'll have the tools to try compressing and creating 3D scenes yourself, opening up fun new possibilities.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 48 to 62 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 scenetok?

SceneTok compresses 3D scenes from multiple views into a compact, unstructured sequence of tokens that you can diffuse like latents in image models. It solves the pain of bulky 3D representations by enabling novel view rendering and scene generation from just a few input images plus poses. This CVPR 2026 project from Max Planck researchers offers a scene autoencoder pipeline, with code releases planned for datasets like RealEstate10K and DL3DV.

Why is it gaining traction?

Unlike dense Gaussian splats or NeRFs that bloat storage, SceneTok delivers a 1D token space that's lightweight and generative, perfect for diffusion-based editing or sampling. Developers dig the CVPR github template vibe—similar to hot cvpr 2024 papers github and cvpr 2025 papers github drops—hyped by its arXiv paper and website demos. Early buzz comes from the promise of chaining image compressors with flow-based decoders for real-world scene tasks.

Who should use this?

Vision researchers prototyping 3D scene gen from sparse photos, like AR/VR devs building immersive environments. Graphics engineers at startups tackling novel view synthesis for virtual tours. Anyone forking cvpr rebuttal github or cvpr poster github setups to extend compressed scenetap workflows.

Verdict

Hold off until code drops—right now it's a polished README with 43 stars and 1.0% credibility score, no runnable models yet. Bookmark for post-CVPR 2026 if diffusable scene tokens fit your stack; otherwise, stick to mature alternatives like Instant-NGP.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.