kenji-tojo

kenji-tojo / diffsoup

Public

DiffSoup: Direct Differentiable Rasterization of Triangle Soup for Extreme Radiance Field Simplification (CVPR 2026)

kenji-tojo.github.iopublicationsdiffsoup
44
1
100% credibility
Found Mar 31, 2026 at 44 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
C++
AI Summary

DiffSoup simplifies complex 3D radiance fields into unstructured triangle meshes with neural textures for fast rendering on consumer devices.

How It Works

1
🔍 Discover DiffSoup

You stumble upon a clever tool that turns detailed 3D scenes into lightweight versions that render smoothly on phones and laptops.

2
📥 Download sample scenes

Grab free 3D datasets like a cozy kitchen or colorful Lego to experiment with right away.

3
⚙️ Prepare your setup

Follow easy steps to get everything ready on your computer so you can start creating.

4
🚀 Simplify a 3D scene

Choose a scene and watch it magically shrink into a fast-rendering model while keeping all the realistic details.

5
👀 Explore in a viewer

Spin around your new 3D model in a simple window, zooming and panning just like a pro.

6
📱 Test on any device

Export to the web or phone to see super-fast playback that feels buttery smooth.

🎉 Share lightweight 3D magic

Now everyone can enjoy your stunning 3D scenes instantly on everyday gadgets without lag.

Sign up to see the full architecture

5 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 diffsoup?

DiffSoup simplifies radiance fields—like those from 3D Gaussian splatting—into unstructured triangle soups with neural textures and binary opacity for extreme model reduction. It uses direct differentiable rasterization in C++ with CUDA, enabling stable optimization and standard depth-tested rendering that runs interactively on laptops or mobiles. Train on datasets like MipNeRF-360 or NeRF-Synthetic via simple Python scripts, then view results in OpenGL, web, or benchmark FPS on phones.

Why is it gaining traction?

Unlike smooth rasterizers that struggle with binary opacity, DiffSoup's stochastic masking delivers stable gradients for aggressive simplification without quality loss. It integrates seamlessly into graphics pipelines for real-time rendering across devices, beating baselines like MobileNeRF in mobile benchmarks. The CVPR 2026 paper and ready-to-run examples hook researchers chasing compact radiance field representations.

Who should use this?

Computer vision researchers optimizing NeRFs for deployment, graphics engineers building mobile AR/VR apps, or 3D reconstruction devs needing sub-15K triangle models from multi-view data. Ideal for anyone evaluating radiance field compression on consumer hardware, with scripts for COLMAP or Blender scenes.

Verdict

Promising for radiance field simplification research, with solid docs, examples, and cross-platform viewers—installs cleanly on Ubuntu/CUDA. Low 44 stars and 1.0% credibility reflect early-stage maturity; test on your datasets before production.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.