Serverless RunPod endpoint that trains a 3D Gaussian Splatting model using Brush, built from source. Part of the RunSplat pipeline.

Input: COLMAP workspace (base64 tar.gz) from colmap-serverless
Output: 3D Gaussian Splat .ply + web-ready .splat binary (base64)

colmap_workspace_b64 (tar.gz)
    │
    ▼  untar → colmap/ directory
    │
    ▼  Brush training (GPU, configurable steps)
    │
    ▼  PLY export (3D Gaussian Splat)
    │
    ▼  convert PLY → SPLAT (web-optimised binary)
    │
    ▼  ply_base64 + splat_base64

3D Gaussian Splatting (3DGS) is a novel view synthesis technique that represents a scene as millions of tiny ellipsoidal "Gaussians", each with a position, orientation, size, opacity, and colour. Unlike traditional meshes or point clouds, the result lets you fly through the scene and see it exactly as it is — from every angle — with real photographic textures.

Why it matters for drone professionals:

• See your land as it really looks — not an abstraction, not a wireframe, but the actual site
• Easier to understand than technical maps — clients and decision-makers get it immediately
• Ideal for design and planning — walk through the future space before breaking ground
• Remote measurements — measure distances, areas, and heights from the browser
• No heavy software — view and share via URL, runs in any modern browser with WebGL
• Track construction or erosion — compare scenes from different flights over time

Explore the ecosystem:

• superspl.at — The Home for 3D Gaussian Splatting
• antimatter15/splat — WebGL Gaussian Splat Viewer (the viewer embedded in RunSplat)

The quality of the Gaussian Splat depends heavily on the diversity of viewing angles during capture. Circular flight patterns dramatically outperform traditional grid surveys:

Unlike traditional grid or double-grid patterns, Circlegrammetry enables drones to fly in circular patterns, with the camera angled between 45° and 70° toward the center of each circle. This method captures images from more angles in fewer flights.

— SPH Engineering, Circlegrammetry

More viewing angles → richer Gaussian initialisation from COLMAP → better-trained splat.

Brush is compiled with lld linker to avoid OOM link failures during cargo build.

{
  "input": {
    "colmap_workspace_b64": "<base64 tar.gz from colmap-serverless>",
    "steps": 30000,
    "eval_every": 1000,
    "export_every": 5000
  }
}

{
  "ply_base64": "<base64 PLY file>",
  "splat_base64": "<base64 SPLAT binary for WebGL>",
  "num_gaussians": 1842311,
  "training_time_seconds": 1187,
  "steps_completed": 30000,
  "status": "done"
}

• Docker with NVIDIA Container Toolkit
• NVIDIA GPU with CUDA 12.9 driver support and ≥24 GB VRAM (for large scenes)
• A COLMAP workspace fixture (see below)

# First build compiles Brush from source — allow ~15–25 min
docker build -t brush-serverless:local .

# Stream output and save log
docker build -t brush-serverless:local . 2>&1 | tee /tmp/build.log

# Test only the Rust compilation stage
docker build --target brush-builder -t brush-builder . 2>&1 | tee /tmp/build-brush.log

The test requires a pre-generated COLMAP workspace fixture (lighthouse.mp4, 30 frames). It is downloaded automatically from the brush-serverless GitHub releases:

./scripts/test_local.sh

# Skip rebuild if already built
./scripts/test_local.sh --no-build

To generate your own fixture from a local video:

# 1. Run colmap-serverless test and capture the workspace
docker run --rm --gpus all colmap-serverless:local \
  python3 handler.py --test_input '{
    "input": {
      "video_url": "https://github.com/alx/runsplat/releases/download/v0.1.5/lighthouse.mp4",
      "num_frames": 30
    }
  }' | python3 -c "
import sys, json, base64
out = json.loads(sys.stdin.read().split('\n')[-2])
open('/tmp/colmap.tar.gz', 'wb').write(base64.b64decode(out['colmap_workspace_b64']))
print('Saved /tmp/colmap.tar.gz')
"

# 2. Use the fixture in the brush test
WORKSPACE_B64=$(base64 -w 0 /tmp/colmap.tar.gz)
docker run --rm --gpus all brush-serverless:local \
  python3 handler.py --test_input "{\"input\":{\"colmap_workspace_b64\":\"$WORKSPACE_B64\",\"steps\":500}}"

1. Verify local build and test pass
2. Push to GitHub and create a release: gh release create v1.0.0 --generate-notes
3. RunPod Hub detects the release, builds the image, runs .runpod/tests.json
4. After tests pass, submit for review on the Hub page

Configuration is in .runpod/hub.json. Available presets: Fast preview (5k), Standard quality (30k), High quality (60k).

• Brush GitHub
• 3DGS original paper — Kerbl et al., SIGGRAPH 2023
• superspl.at — Community hub for Gaussian Splatting
• antimatter15/splat — WebGL viewer used in RunSplat

• Brush — Arthur Brussee, 3DGS trainer
• antimatter15/splat — Kevin Kwok, WebGL viewer
• COLMAP — Schönberger & Frahm, Structure-from-Motion
• RunPod — Serverless GPU infrastructure
• SPH Engineering — Circlegrammetry technique