Gaussian Render Capture – synthetic COLMAP datasets for Gaussian Splatting

A Blender add-on that turns a 3D model into a ready-to-train dataset for Gauss­ian Splat­ting. It places cam­eras on a sphere around your model, ren­ders one image per camera and exports a COLMAP dataset – images, exact camera poses and a start point cloud – for Post­shot and LichtFeld Studio. It started from Gauss Cannon by Arash Kesh­mirian – thanks for the impulse.

A built-in step-by-step guide leads through it; blue but­tons show the next action, red status lines what is still miss­ing.

See the result in your browser – the exam­ple of the tuto­rial, trained from one dataset (320 views) with both train­ers: Post­shot result on SuperSplat · LichtFeld Studio result in the LichtFeld gallery

Example model: LDraw model of set 10252 by Roland Dahl (RolandD), CC BY 2.0 – included in exam­ples/10252_Volkswagen_Beetle.

Documentation

• Tuto­rial – the whole work­flow step by step, with screen­shots
• Quick start – the same on one page
• Ref­er­ence – every set­ting explained
• Trou­bleshoot­ing · Glos­sary · Known issues · Changelog

Highlights

• Guide for begin­ners, one step at a time, with checks.
• Prepare Scene sets up Cycles on the GPU with cap­ture render set­tings.
• Camera sphere – 20 to 1280 cam­eras; Fill Each View and Center in Each View make the model fill and centre every square image.
• Scene ver­sions – <Name>_v001.blend ren­ders and exports into its own dataset folder <Name>_COLMAP/v001/; ver­sions never mix.
• Live Camera Adjust – fine-tune one view or the whole sphere through the camera.
• Render from the panel – PNG or TIFF; Render EEVEE (fast) or Render Cycles (exact); or Command Line Render: a double-click script ren­ders head­less while Blender stays free.
• Start point cloud from all model ver­tices plus sur­face points, in the mate­rial colours; a GPU vis­i­bil­ity filter drops hidden inner parts.
• Auto scale with the factor recorded in <vNNN>_gcapture_export.json next to the dataset, so a trained splat can be placed back onto the model.
• Clean Splat – after train­ing, removes floaters from the trained splat with the model's depth per camera: no ren­der­ing, a few sec­onds, the orig­i­nal file stays.

Install

Blender 4.1 or later. Tested by hand on Windows 11 (4.1, 4.4, 4.5, 5.0, 5.2, 5.3). On Linux (Ubuntu) and macOS (Apple Silicon) only the auto­mated test suite has run, with Blender 4.1 and 5.2, on GitHub's vir­tual test machines – not yet on phys­i­cal com­put­ers. GPU ren­der­ing on AMD, Intel and Apple GPUs has not been tried yet – reports are wel­come in the issues.

For train­ing: Post­shot runs on Windows, LichtFeld Studio on Windows and Linux with an NVIDIA GPU. On macOS use a trainer that reads COLMAP datasets.

⬇ Latest release – down­load gaussian_capture-<version>.zip there; do not unpack it.

1. Edit → Pref­er­ences → Add-ons → Install from Disk… and pick the down­loaded ZIP. Blender 4.2 and later install it as an exten­sion, Blender 4.1 as an add-on.
2. Enable Gauss­ian Render Capture (dis­able any older ver­sion first).
3. The panel is in the 3D view­port side­bar (N), tab Gauss­ian Render Capture – press Start Guide.

Updat­ing – restart Blender after installing a new ver­sion. Up to 1.3.0 the add-on was a single file: on Blender 4.2 and later first remove the old Gauss­ian Render Capture (gaussian_capture.py) in Pref­er­ences → Add-ons, then install the ZIP – enabling the new ver­sion next to the old one is refused with a mes­sage. On Blender 4.1 installing the ZIP replaces the old file. Your scenes keep all set­tings.

Upgrad­ing from Gauss­ian Render Scan (1.0.x) – the add-on was renamed in 1.1.0. Disable and remove Gauss­ian Render Scan, then install the ZIP. Scenes made with it are taken over auto­mat­i­cally when you open them (set­tings, camera sphere, live adjust­ments); save them to keep the new names.

Related projects

• Gauss Cannon by Arash Kesh­mirian (Warp­gate Labs) – the add-on this one started from. Camera paths from helper meshes, a ray-traced coloured point cloud and camera export for LichtFeld Studio and Post­shot.

Licence and credits

Copy­right (C) 2026 Prof. Michael Klein; por­tions (ray-cast­ing helpers) Copy­right (C) 2025 Arash Kesh­mirian / Warp­gate Labs. Licensed under GPL-3.0-or-later (see LICENSE).

Started from Gauss Cannon by Arash Kesh­mirian (Warp­gate Labs, GPL-3.0-or-later); today only its ray-cast­ing helpers remain (inte­rior-camera test, CPU fall­back of the vis­i­bil­ity filter). Details and all other third-party notices: THIRD_PARTY.md.

Prof. Michael Klein
Digital Film Design – Ani­ma­tion/VFX · Medi­ade­sign Uni­ver­sity of Applied Sci­ences
medi­ade­sign.de · vir­tu­al­re­pub­lic.org · ren­der­bricks.com

Trans­parency note on the use of AI: the author is not a pro­gram­mer but has worked in CGI since 1987. The add-on was devel­oped entirely through vibe coding with Anthropic Claude (Claude Code): Claude wrote the code, the tests and this doc­u­men­ta­tion from his descrip­tions. The con­cept, the design deci­sions, the tests in Blender, Post­shot and LichtFeld Studio and the accep­tance of every ver­sion are the author's; he is respon­si­ble for the con­tent.

LEGO® is a trade­mark of the LEGO Group; this and all other trade­marks named here belong to their owners, none of whom spon­sors, autho­rises or endorses this project (see THIRD_PARTY.md).