Automatic reconstruction
Choose a video or image folder and run the full pipeline with a preset. You can also run each stage separately.
Video-to-3D Gaussian Splatting software
Rebuild reality.
Import a video or a folder of photos. Reconstrura estimates the camera positions, reconstructs the scene, and trains a Gaussian Splat on your NVIDIA GPU. Export the result as a .ply file. Everything stays on your computer.
Free trial · One-time $25 · Windows and Linux · NVIDIA GPU with 8 GB+ VRAM
- Processing runs on your machine
- Your footage is never uploaded
- Windows & Linux
- NVIDIA GPU with 8 GB+ VRAM
Features
From source footage to a finished splat
Frame extraction, masking, camera reconstruction, training, and export in one desktop app.
GPU-accelerated processing
Feature extraction, matching, bundle adjustment, and splat training use your NVIDIA GPU.
Support for 8 GB VRAM
NVIDIA GPUs with 8 GB of VRAM are supported. Reconstrura adjusts the Gaussian budget to the available memory.
Local processing
Videos and photos are processed on your computer and are not uploaded. Reconstruction works offline after the required components are installed.
Video, photos, and 360° footage
Import common video formats, a folder of photos, or equirectangular 360° footage. Reconstrura converts 360° footage into a virtual camera rig.
Frame selection
Reconstrura scores frames for sharpness and spacing, removes near-duplicates, and keeps useful viewpoints.
Optional masking
Use SAM 2.1 or EdgeTAM to exclude people, sky, or background regions. Review each detected instance before reconstruction.
Live previews
Inspect the point cloud during reconstruction and the splat during training. Both viewers can be orbited while processing continues.
Presets and manual controls
Start with a preset or adjust the feature count, resolution, matching, Gaussian limit, iteration count, and other stage settings.
Standard output formats
Export a .ply file for SuperSplat, Postshot, or Brush, or an undistorted COLMAP dataset with transforms.json.
How it works
How a project is processed
This example uses a 41-second phone video of a rose. It produced 125 frames, 125 solved camera positions, and one Gaussian Splat.
Before you start
Welcome
Create a project or reopen an existing one. Reconstrura records the last completed stage so you can continue where you stopped. The background shows the point cloud from a completed project.
project folders · resume at any stage · live point-cloud backdrop
Background: a 514-image rhododendron capture with 514 cameras and 226,311 points
Step 1 of 5 · Source
Import
Choose a video or a folder of photos. Reconstrura displays the resolution, frame rate, codec, and duration. Existing projects can continue from any completed stage.
MP4 / MOV / MKV / INSV · photo folders · 360° equirectangular → virtual camera rig
Example project: yellow-rose · 41-second phone video · 2160×3840 · 30.06 fps
Step 2 of 5 · Frames
Extraction
Reconstrura scores frames for sharpness and spacing. It removes near-duplicates while keeping enough different viewpoints to reconstruct the scene.
sharpness scoring · redundancy filtering · writes images/
scanned 1,257 candidates · kept 125 keyframes · sharpness-aware selection
Step 3 of 5 · Clean-up
MaskingOptional
Choose whether to exclude people, sky, or background. Reconstrura lists each detected instance so you can include or exclude it without running segmentation again.
EdgeTAM · SAM 2.1 S / L · per-instance selection · writes masks/
Example: a passer-by is excluded from 14 of 125 frames · red areas are ignored during reconstruction
Step 4 of 5 · Cameras
Reconstruct
Reconstrura estimates the camera position for each frame and builds a sparse point cloud. Masked areas are ignored. You can inspect the point cloud during and after reconstruction.
ALIKED + LightGlue · GPU bundle adjustment · free-space carving · writes sparse/0
Solved example: 125 cameras, 108,212 points, and 1.18 px reprojection error in 2m 02s
Step 5 of 5 · Splats
Train
Reconstrura trains a Gaussian Splat from the solved cameras. You can inspect the live result during training and export the final .ply file.
3DGS · VRAM-aware gaussian budget · adaptive growth stop · writes splat/point_cloud.ply
Example result: 40,000 iterations in 13m 35s · 344,182 Gaussians · 32.91 dB PSNR · 2.25 GB peak VRAM
Project files are saved after each stage. Frames are written to images/, masks to masks/, the COLMAP model to sparse/0/, and the final splat to splat/point_cloud.ply. To export an undistorted dataset with transforms.json, run reconstrura --export-undistorted.
Pricing
$25 one-time purchase
The license includes future updates and use on up to three devices at once. There is no subscription or per-scene charge.
One-time purchase
$25
VAT included
- Future updates included
- Use on up to three devices at once
- Free trial without an account or payment card
- Windows & Linux
- Footage processed locally
Sold by Polar (Merchant of Record) · Licensed by Elias Rehfus under the EULA.
Download
Get Reconstrura
Install the full app, then start the free trial when you are ready. No account or payment card is required.
Linux AppImage
For current Linux distributions. Make the file executable and run it. In-place updates are supported.
Download AppImageLinux .deb
For Ubuntu 22.04 and newer, and other Debian-based distributions. Includes the CLI and supports updates through APT.
Download .deb
Recommended for Linux
Flatpak
A sandboxed installation for most Linux distributions. Updates are managed by your software centre.
Install FlatpakSHA-256 checksums for version 1.1.0: https://reconstrura.com/dl/1.1.0/SHA256SUMS
Reconstrura uses Qt, FFmpeg and libheif under the LGPL. License terms, exact source code, and library replacement instructions are available without an account.
System requirements
GPU
NVIDIA, 8 GB+ VRAM (GTX 16xx / RTX 20xx or newer)
Driver
NVIDIA 525+ (580+ recommended)
RAM
16 GB recommended
Disk
~10 GB free (one-time runtime download)
OS
Windows 10/11 (64-bit) · Linux (Ubuntu 22.04+, Fedora, Arch)
FAQ
Questions
What is Gaussian Splatting?
Gaussian Splatting represents a 3D scene with many small, coloured Gaussian primitives. It supports real-time rendering and can preserve fine detail and soft edges. Reconstrura estimates the camera positions from your images, then trains the splat on your GPU.
Do I need an expensive PC?
You need an NVIDIA GPU with at least 8 GB of VRAM. A faster GPU reduces processing time, while more VRAM allows a larger Gaussian budget.
Does it work offline?
Reconstruction runs locally and does not require an internet connection. On first launch, Reconstrura downloads the GPU runtime and core models. The download is about 2.4–4.0 GB and uses roughly 6–7 GB after installation. Masking models are downloaded when masking is first used. Starting a trial requires a one-time request to the license server. Purchased licenses are revalidated periodically, and the app also checks for updates.
What does the trial send?
When you start the trial, Reconstrura sends a pseudonymized hash of the device identifier, along with the platform and app version. The trial request does not send your name, email address, usage statistics, photos, or videos. Separate privacy-safe product analytics is optional, off by default, and described before you enable it in Settings. See the privacy policy for details.
How many devices can I use?
A license can be active on up to three devices. Activating a fourth device deactivates the least recently used one. You can also release a device from the app’s settings.
What about refunds?
Polar processes purchases and refunds. Use the Request a refund form to submit a request and receive a timestamped receipt. Eligible requests with verified order details can be refunded automatically. Other requests are sent to Reconstrura support for review. This does not affect your statutory rights for defective digital products.
Which GPUs and operating systems are supported?
You need an NVIDIA GPU with at least 8 GB of VRAM and driver version 525 or newer. Reconstrura supports 64-bit Windows 10 and 11. Linux packages are available as an AppImage, a Debian package, and a Flatpak. AMD GPUs and Apple Silicon are not supported.
Do updates cost extra?
No. The one purchase covers every future version. There are no paid upgrades and no subscription.
Changelog
Release notes
1.1.0 · Tracked-scene imports, mask sequences and smarter extraction
Bring existing COLMAP, NeRF / Instant-NGP and OpenUSD scenes directly into Reconstrura, import mask sequences from other tools, and extract a smaller, better-connected set of frames.