Step-by-step workflow: multi-view photos → COLMAP camera poses → Brush Web GSplat training → export and view .ply.

gsplat-from-multiview-workflow.md

GSplat from Multi‑View Images

Photos → COLMAP (poses) → Brush (GSplat) → View

Requisites

• COLMAP: https://github.com/colmap/colmap/releases
• Chrome / Edge (WebGPU) — required for Brush Web Demo: https://arthurbrussee.github.io/brush-demo/
• Images: .jpg or .png, static scene, sharp, consistent exposure and white balance, 60–80% overlap

---

1. Prepare

<project>/
  images/
    *.jpg | *.png

(All photos only in images/)

2. COLMAP (Prepare Camera Pose Dataset)

1. File → New project

   • Database: <project>/<name>.db
   • Images: <project>/images/
   • Click Save

2. Processing → Feature extraction → Extract

3. Processing → Feature matching → Run

4. Reconstruction → Start reconstruction

5. File → Export model → <project>/ → This step creates cameras.bin, images.bin, points3D.bin

6. (Optional) File → Export model as… → .ply → <project>/

3. Brush (Train GSplat)

Open: https://arthurbrussee.github.io/brush-demo/

1. File → select COLMAP .ply
2. Directory → select <project>/ (where .bin files and images/ exist)
3. Start / Train → wait until stable
4. Export → PLY → GSplat file

Output example: <scene>_gsplat.ply

4. View Online

Open: https://superspl.at/editor

• Drag & drop GSplat .ply file
• Orbit / zoom, adjust size and opacity

These GSplats were generated using the workflow described in this guide:

• Desk Workspace
• Toy Bird

The HTML viewers were exported using SuperSplat.