Frankensplat trains splats in your browser

Rafael Spring’s Frankensplat brings local Gaussian-splat training to the browser with JavaScript and WebGPU, using COLMAP reconstructions as input.
Frankensplat browser interface for training Gaussian splats with WebGPU
Screenshot: OfflineTools / Rafael Spring, via Radiance Fields.

For those who don’t know the tool: Frankensplat trains Gaussian splats from COLMAP camera reconstructions directly in a browser using WebGPU. No rented render farm, no installation ceremony, just a graphics card being asked to do the actual work.

Frankensplat web interface for Gaussian splats showing photo upload and COLMAP reconstruction options with dark theme.

Gaussian splats have found another surprisingly lively home: a browser tab. Spring’s new companion to browser-based photogrammetry handles the training stage. The current workflow still needs a COLMAP reconstruction first. It is a neat division of labour: reconstruct the cameras first, then let WebGPU make the splat. Now that is one way to make a splash.

COLMAP first, splats next

The required SfM reconstruction can come from Spring’s browser-based COLMAP tool, which we covered in COLMAP Moves Into the Browser. That earlier stage reconstructs camera information and sparse scene structure; Frankensplat then handles Gaussian-splat training. The product overview introduces the new tool.

Spring has also updated the COLMAP implementation. He reports feature detection running about 50 per cent faster, matching around three times faster and SfM about 30 per cent faster. His announcement does not specify the test datasets, hardware or baseline versions for those improvements.

Fast, with the benchmark attached

The developer’s study compares 26 completed runs on an RTX 3070 with 8 GiB VRAM under Windows: Frankensplat in Edge/WebGPU and a development build of LichtFeld Studio in native CUDA. Runs used 15,000 iterations and a 1.5-million-splat cap. Fixed-resolution Frankensplat took 30.3 per cent less time across five matched scenes than LFS with detail features disabled. Across seven scenes, progressive Frankensplat reached mean fitted PSNR of 28.52 dB versus 29.39 dB for LFS detail-off. These are single runs with different growth schedules; the appearance-fitted metric is not an untouched held-out score. This is a vendor comparison, not an independently reproduced speed guarantee.

A combined, automatic browser workflow from photographs to splats is a possible future direction. Spring says the two projects may eventually merge. The guide accepts folders or ZIPs containing images and a COLMAP sparse reconstruction;

Feature detection software interface analyzing multiple images of stone birdbath with keypoints for visual anchors.

PLY is supported for viewing and export. Training requires WebGPU and sufficient GPU limits and memory; viewing uses WebGL2. Controls include splat budget, image scale and PPISP appearance correction, with progressive resolution enabled by default. Checkpoints resume against the matching original dataset. The site and modules must first load online; subsequent offline access depends on browser caching.

Availability

Spring has announced Frankensplat as released and free to use. The application, product overview and study are linked below.

A browser tab can now be a splat workshop. That is genuinely exciting, provided nobody closes it at the dramatic moment. Test the tool on representative capture sets and target hardware, and verify the resulting assets in the intended downstream applications before using it in production.

https://offlinetools.io/frankensplat

ProductFrankensplat
DeveloperRafael Spring
InputImages and a COLMAP Structure-from-Motion reconstruction
ProcessingLocal browser-based training using JavaScript and WebGPU, as described by the developer
OutputGaussian splat asset in PLY format
AvailabilityFree browser tool, according to the developer