IrieVybz lab · /splat

The post is selling 3D Gaussian Splatting as a free real-estate product. The repo is a 2023 research trainer.

Fran Pradas pointed at this thread. The GitHub link with 23k stars is graphdeco-inria/gaussian-splatting, the official implementation of Kerbl et al. It is the origin of the technique — not the version we should ship.

What the tweet claims

25 Aug 2026 · Fran Pradas

Turns a house into a 3D tour from photos

Called 3D Gaussian Splatting, and it is free

No app, no VR headset, no appointment

People charge $300–$800 for scans

The GitHub repo has 23,000 stars

Loading a real photographed kitchen…

Real capture

Kitchen

Mip-NeRF 360 · real photographs

Drag to orbit · right-drag to pan · scroll to zoom

Do not train with the starred repo

Inria and MPII license that code for research and evaluation only. Commercial use — including paid property tours — needs written consent from stip-sophia.transfert@inria.fr. The tweet calling it “gratis” skips that.

It is not a one-click tour

You need overlapping photos or video, camera poses, an NVIDIA GPU (or a paid capture app), 15–45 minutes of training per space, floater cleanup, and a web host for a 10–80 MB scene. No VR is true. “No app” is marketing.

The technique is still worth owning

Photoreal walkthroughs in a browser are real. $300–$800 scan fees exist. We add the capability with Apache/MIT tools, not a fork of the INRIA trainer.

How it works

Five stages, not one magic binary

1. Capture

Walk the space slowly. 150–300 stills or a steady phone video, 60–70% overlap, every room and doorway. Miss a wall and the reconstruction tears.

2. Pose the cameras

COLMAP (or a capture app) estimates where each photo was taken and builds a sparse point cloud. No poses, no splat.

3. Grow Gaussians

Each point becomes a 3D ellipsoid with position, scale, rotation, opacity, and view-dependent color. Training compares rendered views to the photos and splits or prunes Gaussians until the scene matches.

4. Splat, do not ray-march

At view time the ellipsoids are projected onto the image and alpha-composited. That is why this is real-time in a browser while classic NeRF is not.

5. Clean and ship

Raw PLY files are 100–300 MB per room. SuperSplat removes floaters, then you compress to .splat / .ksplat / .spz and embed a viewer.

What is in that repo

Four pieces of research code

train.py — PyTorch optimizer. Wants CUDA 11/12, a compiler, and ideally 8–24 GB of VRAM. Default run is 30,000 iterations.
convert.py — wraps COLMAP so a folder of images becomes posed cameras plus a sparse cloud.
diff-gaussian-rasterization — custom CUDA kernels that make training and the SIBR viewer real-time.
SIBR viewers — native OpenGL apps, not a web tour. Web viewing is a later ecosystem (this desk, SuperSplat, antimatter15).

Better version

Use gsplat. Keep SuperSplat. Ignore the star count.

Nerfstudio's gsplat reimplements the method from scratch under Apache 2.0. It trains in less time, uses less memory, and is the library every serious 2026 pipeline sits on. SuperSplat is the editor. Phone apps cover capture when we do not want a GPU box. The INRIA repo remains useful as a paper artifact and a quality baseline — not as production code.

gsplat SuperSplat

Arsenal decision

Tool	Job	License	Call
INRIA 3DGS	Paper reference	Research only	Highest citation count. Wrong default for Nexus work.
gsplat + Nerfstudio	Train locally	Apache 2.0	Faster, less VRAM, commercially usable. This is the training stack.
SuperSplat	Edit	MIT	Browser cleanup, crop, and export. Use after every training run.
Polycam / Scaniverse	No-GPU capture	App terms	Phone to splat when there is no RTX box. Pay for export on some tiers.
This desk	View + brief	MIT viewer	Spark viewer on real captures. Do not train with the INRIA repo.

The Cursor skill in .cursor/skills/3d-gaussian-splatting/SKILL.md is the operational version of this page: capture rules, the gsplat train path, SuperSplat cleanup, and how to drop the result into this viewer.

Boardroom check

Tech and commercial seats reviewed whether we should chase a better primitive than 3DGS. Short answer: keep Gaussians, leave the 2023 repo, and do not treat our shader as the product.

Tech

Hybrid. Train with gsplat (Apache). Deliver with SuperSplat + SOG/SPZ. View in Spark or this desk. Skip Hierarchical 3DGS, Scaffold-GS, and SuGaR — paper repos that inherit the Inria license.

Commercial

Own-the-renderer is the wrong bet. Fastest cash is phone 360 (Kuula / CloudPano). Measure and floor plans still come from Matterport, iGUIDE, or Polycam Space — not from a splat.

Call

Partner and resell tours. Keep Splat Desk as the lab and skill. Build a listing desk later (MLS URL, embed, CTA) only if a client is paying per room.