ziplab

ziplab / TriSplat

Public

TriSplat: Simulation-Ready Feed-Forward 3D Scene Reconstruction

lhmd.toptrisplat feed-forward-reconstruction novel-view-synthesis triangle-splattiing
90
0
94% credibility
Found May 26, 2026 at 90 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

TriSplat is an academic research project that transforms multiple unposed photographs into complete 3D scene reconstructions. Unlike traditional approaches that require known camera positions or need multiple processing steps, TriSplat predicts camera poses and generates simulation-ready triangle meshes in a single forward pass. The system was developed by researchers at Zhejiang University and ETH Zurich, trained on large video datasets, and outputs meshes compatible with standard 3D graphics and simulation software.

How It Works

1
📸 You have photos of a space

You took several pictures of a room, building, or outdoor scene from different angles, but you don't know exactly where each photo was taken.

2
🧠 Your photos become a 3D model

TriSplat looks at your photos and figures out both where they were taken and what the 3D shape of everything looks like, all in one step.

3
✨ You get a complete 3D mesh

Unlike other tools that give you a messy point cloud or require extra steps, TriSplat delivers a clean triangle mesh you can use immediately.

4
🎮 Your mesh works in games and simulators

The exported 3D model opens directly in popular tools like Blender, Unity, or physics simulators without any conversion needed.

🎉 Your scene is ready for anything

You now have a fully usable 3D reconstruction that can be animated, simulated, or integrated into any project you have in mind.

Sign up to see the full architecture

3 more

Sign Up Free

Star Growth

See how this repo grew from 90 to 90 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 TriSplat?

TriSplat is a Python project that reconstructs complete 3D scenes from just a handful of unposed images. Unlike traditional approaches that output messy point clouds or require slow per-scene optimization, TriSplat runs as a single feed-forward pass and produces clean triangle meshes ready for physics simulation. It predicts geometry, camera poses, and appearance attributes simultaneously, then exports standard mesh formats that load directly into tools like Blender, Isaac Sim, Unity, or PyBullet. The pipeline uses PyTorch with CUDA acceleration and supports datasets like RealEstate10K, DL3DV, and ScanNet.

Why is it gaining traction?

The killer feature is speed combined with usability. Traditional NeRF and Gaussian Splatting pipelines require minutes to hours of optimization per scene and then need messy post-processing to extract meshes. TriSplat skips the optimization entirely and outputs simulation-ready geometry in one shot. Researchers and developers building scene reconstruction pipelines are drawn to the fact that the output just works in standard graphics and simulation tools without conversion steps. The pretrained checkpoints on HuggingFace lower the barrier to experimentation.

Who should use this?

Game developers and robotics researchers who need fast scene reconstruction for asset generation or environment digitization will find this most useful. If you're building a pipeline that processes many scenes and can't afford per-scene optimization time, TriSplat solves that problem. Academic researchers in 3D vision working on feed-forward reconstruction benchmarks will appreciate the reproducible training setup with Hydra configs and Lightning. However, if you need the absolute highest reconstruction quality with unlimited compute time, traditional optimization-based methods still win.

Verdict

TriSplat solves a real pain point in the 3D reconstruction workflow, but the project is early-stage with only 90 stars and a credibility score of 0.95%. The documentation is functional but sparse on production deployment guidance. Worth exploring for prototyping and research, but plan to invest time in understanding the configuration system before relying on it for critical pipelines.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.