Skip to main content

SkySplat

ADDON
📖4 min read
Latest Version: 0.4.2
SkySplat (free): Gaussian Splatting in Blender made easy by Nicko16
1 / 2

Introduction​

SkySplat is a Blender addon that handles the full pipeline for creating 3D Gaussian Splats from drone footage: video import, frame extraction, COLMAP reconstruction, model transformation, camera animation, and Gaussian Splatting training, all within Blender. Multi-instance support lets you process several flights of a site in one .blend without file collisions. It serves investigators reconstructing sites from drone/aerial footage who want one Blender-native pipeline instead of stitching together external SfM and training tools.

Key Takeaways​

Full drone-to-splat pipeline

Video import, frame extraction, COLMAP, model alignment, cameras, and training, all inside Blender.

Multi-instance, no collisions

Several datasets per .blend with independent settings and paths. Ideal for multi-pass drone coverage.

Free + open source

Blender 4.0+ (fully supported in 5.0). Camera paths double as a verification aid.

Installation​

  1. Download from the GitHub repo (free, open source).
  2. Install in Blender via Preferences → Add-ons → Install, selecting the downloaded archive.
  3. Dependencies: COLMAP (for reconstruction) and the Brush App binaries, which ship with the releases.

Supports Blender 4.0 and later (fully supported in 5.0).

Usage​

Import and extract frames· footage

Load one or more videos with independent settings, set start/end/step frame extraction, and let the addon create per-video output folders. SRT metadata files are detected and loaded automatically; they improve COLMAP initialization.

COLMAP reconstruction· SfM

Run COLMAP per instance (sequential or exhaustive matching). Each instance keeps independent input/output paths. No conflicts when processing multiple reconstructions.

Transform and align· model

Load COLMAP sparse reconstructions into Blender; scale, rotate, and translate to align with the scene. Export for training.

Create the camera path· cameras

Generate animated cameras from COLMAP positions. Keyframes align to each reconstructed frame, resolution matched, with smooth quaternion rotation for flythrough animations.

Train the Gaussian Splat· Brush

Train per-instance with the integrated Brush tools (real-time viewer option), then export your splat.

In an Investigation​

SkySplat's value in an investigation is multi-instance drone coverage: several flights / passes of a site can be processed in one .blend with independent settings, then merged into site context. The animated camera paths double as a verification aid. Overlay them on the source footage to check reconstruction alignment.

What to document when using SkySplat

In Your Method Section​

Document the source footage, frame extraction settings (interval/count), and the COLMAP reconstruction parameters (sequential vs exhaustive). Mention the versions of SkySplat and Blender used.

In Your Decision Log​

  • Parameter choices for frame extraction (why specific frames were chosen).
  • COLMAP matching strategy selected.
  • Gaussian Splatting training iterations and quality settings.
  • Date performed.

Verification​

Overlay the generated camera path with the original video footage to ensure alignment; check the point cloud or splat against known reference points if available.

Example Methodology Statement​

"3D reconstruction was performed using the SkySplat addon (v0.4.0) in Blender 5.0. Frames were extracted from the drone footage at 1-second intervals. COLMAP structure-from-motion was executed using sequential matching. The resulting sparse point cloud was aligned to the ground plane, and a 3D Gaussian Splat model was trained for 30000 iterations using the integrated Brush tools."

Limitations​

  • SfM/splatting quality depends on the input footage. Lighting and motion blur degrade results.
  • Requires COLMAP + Brush App correctly installed. Processing large datasets is time-intensive.
  • Geometric guarantees are weaker than measured photogrammetry. Use for context and visualization, not for dimension-critical forensic claims.
  • The output is model-scale, not georeferenced. If the location must be stated in a report, align the reconstruction with a GPS or GPS-EXIF workflow first.

Resources​