Skip to main content

Spirula Studio

TOOL
📖5 min read
Latest Version: 2026.9.24

Introduction​

Spirula Studio trains 3D Gaussian Splatting models from raw photos or video. The pipeline runs from frames to camera poses to a splat to a textured mesh, all inside one binary. The tool runs outside Blender on Windows, Linux, and macOS. It serves teams that want a full splat pipeline without a Python environment, including machines with AMD or Intel GPUs.

Key Takeaways​

One Binary, Whole Pipeline

Frame extraction, masking, camera solving, training, and meshing live in one application, with no Python stack and no separate COLMAP install.

Trains on Any GPU

Vulkan compute runs on NVIDIA, AMD, Intel, and Apple hardware, and 8 GB of VRAM holds 10 million Gaussians at full spherical harmonics.

Verify the Scale

Metric scale comes from capture telemetry. Check one known distance before you measure anything.

Installation​

  • Download the build for your platform from the GitHub Releases page. Arch Linux users can install the AUR package. The license is GPL-3.0 and the price is zero.
  • Choose a backend at first launch: Vulkan (recommended, runs on NVIDIA, AMD, Intel, and Apple Silicon) or CUDA (NVIDIA GPUs on Windows and Linux).
  • Frame extraction calls ffmpeg by default. Install it on your system, or build from source with GPU decoding, which shifts AVC/HEVC patent compliance onto you.
  • AI masking downloads a SAM checkpoint on first use. The app shows Meta's license terms before it fetches anything.
  • Plan for the VRAM: the trainer fits 10 million full-SH Gaussians into 8 GB.

Usage​

Build the dataset· photos or video

Load the capture. The tool extracts frames from video, runs AI masking, and solves the camera poses with its built-in SfM. Telemetry in the file sets metric scale and orientation.

Train· Vulkan or CUDA

Start training on your GPU. Quantized training keeps the memory footprint low, and the strategy combines MCMC, IGS+, and MRNF sampling to reduce floaters.

Edit· v2026.9.24

Clean the result inside the app. The current release edits Gaussian splats, sparse reconstructions, meshes, and masks.

Mesh

Generate a textured mesh from the trained splat when you need surface geometry instead of a point-based scene.

Export· splat, mesh, renders

Export the trained splat, the textured mesh, and image or video renders of the scene.

Automate· optional

Version 2026.9.24 ships an MCP server, so a connected agent can drive the pipeline for you.

Into Blender​

  • Two outputs cross over: the trained splat and the textured mesh. The renders stay outside the scene as presentation media.
  • The mesh imports through Blender's standard model importers. The splat follows the import path on the Gaussian Splatting page.
  • Metric scale arrives from the capture telemetry since the September 2026 release. Verify one known distance right after import, then trust the dimensions.
  • The output carries no coordinate reference system. Georeference against your baseline in QGIS or in Blender before you align it to other evidence.

In an Investigation​

How Spirula Studio output enters a documented record

In Your Method Section​

  • Record the Spirula Studio version, the backend (Vulkan or CUDA), the source media, the frame extraction and masking settings, the SfM settings, and the export you used.

In Your Decision Log​

  • Record why you chose Spirula Studio over another trainer: cross-vendor hardware, the self-contained pipeline, or the 360° camera support. Record the backend choice and any scale or quality settings.

Verification​

  • Keep the source media and the dataset configuration with the case record. Re-run the training on the same inputs and compare the exports. Store the exported splat, mesh, and renders with file hashes, and confirm the scale against a known measurement.

Example Methodology Statement​

"Frames were extracted from [video file] and camera poses were solved with the built-in SfM module of Spirula Studio [version] on the [Vulkan/CUDA] backend. Metric scale was taken from the capture telemetry and verified against [known distance]. The trained splat and the textured mesh were exported on [date] and stored with SHA-256 hashes in [location]. Measurements in the reconstruction refer to the verified scale."

Limitations​

  • ffmpeg is required for frame extraction in a default install. The GPU-decoding build option puts AVC/HEVC patent responsibility on you.
  • AI masking depends on SAM checkpoints under Meta's licenses: SAM 2.1 is Apache-2.0, and SAM 3 uses Meta's own non-standard terms.
  • The CUDA backend only supports NVIDIA GPUs. macOS and AMD or Intel hardware run on the Vulkan backend.
  • Metric scale only exists when the capture carries telemetry or you supply a known measurement. Still images without metadata arrive unscaled.
  • The release cadence is fast, with several updates in August and September 2026 alone. Pin the exact version you validated for a case record.

How It Works​

The built-in SfM module first solves camera positions from your frames. The trainer then represents the scene as millions of anisotropic Gaussians and optimizes them against the training views with quantized storage, which is what keeps VRAM demand low. Meshing later extracts a surface mesh from that trained splat.

This matters for a report because the splat is an optimized fit to your photos, and it inherits its scale and orientation from telemetry or from your settings. Floaters and blur mark areas the cameras observed weakly. State the settings you used and treat every derived measurement as an estimate that a reader can re-run.

Sources​

  1. Spirula Studio README on GitHub : features, backends, requirements, news log
  2. Spirula Studio v2026.9.24 release on Radiance Fields : editing, render export, MCP server, GPL-3.0
  3. Spirula Studio package on the AUR : Arch Linux install path