Gaussian Splatting

1. git clone https://github.com/graphdeco-inria/gaussian-splatting (the original 3D Gaussian Splatting implementation).
2. pip install -r requirements.txt + CUDA toolkit (needs NVIDIA GPU).
3. Capture a scene: take 50-200 photos of a static scene from different angles (use a phone, orbit around the subject).
4. Run COLMAP first (Structure from Motion, included in the pipeline): python convert.py -s dataset/input_images/ — this estimates camera positions for each photo.
5. Train the Gaussian Splats: python train.py -s dataset/input_images/ -m output/scene_name — training takes 10-60 minutes on RTX GPU for 200 photos.
6. View the result: python viewers/bin/SIBR_viewer.AppImage -m output/scene_name — real-time photorealistic rendering. You can fly through the scene.
7. Alternative (easier): pip install nerfstudio → ns-train gaussian-splatting --data dataset/input_images/ — simplified training, web viewer built-in.
8. Export to standard formats: .ply files can be imported into Unreal, Unity, Blender (with plugins).

Used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb). Trains Gaussian Splatting on 100-200 photos in 20-40 minutes. Real-time rendering at 60+ fps after training. For small scenes (single room, object): training fits in 12 GB. For large scenes (building exterior, 500+ photos): 12 GB may OOM — reduce photo resolution or scene bounds. Pair with Ryzen 5 5600 + 32 GB DDR4 + 1TB NVMe (COLMAP preprocessing is CPU-heavy). Total: ~$390-440. Gaussian Splatting at $400 is viable for small-to-medium scenes.

Used RTX 3090 24 GB ($700-900, see /hardware/rtx-3090). Trains on 500+ photos of large scenes in 30-60 minutes. The 24 GB VRAM handles high-resolution splats (10M+ Gaussians) for professional VFX, architectural visualization, and VR environments. For production photogrammetry pipelines (capturing entire buildings, film sets): RTX 4090 24 GB ($1,600, see /hardware/rtx-4090) reduces training to 15-30 minutes. Total: ~$1,800-2,500. For real-time VR rendering of splats: RTX 4090 at 90 fps for 5M+ Gaussian scenes.

The mistake: Taking 20 smartphone photos while walking in a straight line past a building, running COLMAP, and getting an error: "failed to register enough images." Why it fails: COLMAP needs sufficient overlap and angular diversity between photos. Walking in a straight line provides only forward-facing views with small perspective changes. The camera tracker can't triangulate 3D points from near-identical viewpoints. The fix: Orbit around the subject. Take photos from different heights (knee-level, eye-level, overhead). Ensure 60-80% overlap between adjacent photos. For a room: walk a circle, then walk a smaller circle. For an object: orbit 360° at 2-3 different heights. COLMAP needs parallax — the apparent shift of objects when viewed from different angles. A straight-line walk provides almost zero parallax for distant objects. Orbit = parallax = successful reconstruction.