NeRF (Neural Radiance Fields)

1. pip install nerfstudio (the standard open-source NeRF training framework — wraps multiple NeRF variants).
2. Capture a scene: 50-200 photos of a static scene from different angles. Use a phone camera. Orbit around the subject while keeping 60-80% overlap.
3. Run COLMAP for camera pose estimation: ns-process-data images --data dataset/input_images/ --output-dir dataset/processed/ (processes photos, runs COLMAP, outputs transforms.json).
4. Train a NeRF: ns-train nerfacto --data dataset/processed/ (Nerfacto is the default fast NeRF variant). Training takes 10-30 minutes on RTX GPU for 100 photos.
5. View the result: ns-viewer --load-config outputs/processed/nerfacto/config.yml → opens a web viewer at localhost:6006. Navigate the 3D scene in real-time.
6. Render novel views: ns-render camera-path --load-config outputs/processed/nerfacto/config.yml --camera-path-filename camera_path.json --output-path renders/
7. Reality: Gaussian Splatting (see /tasks/gaussian-splatting) has largely superseded NeRF for real-time rendering. NeRF is still relevant for research, volumetric effects (fog, smoke), and scenarios where splats struggle (transparent objects, reflections).

Used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb). Trains Nerfacto on 100 photos in 15-30 minutes. Renders novel views at 5-10 fps after training (not real-time — this is why Gaussian Splatting won). For research/learning: perfectly usable. For production real-time rendering: use Gaussian Splatting instead. Pair with Ryzen 5 5600 + 32 GB DDR4 + 1TB NVMe. Total: ~$390-440. NeRF training is more VRAM-efficient than Gaussian Splatting (NeRF is a small MLP; splats store millions of explicit Gaussians). A 12 GB card handles larger scenes in NeRF than in GS.

Used RTX 3090 24 GB ($700-900, see /hardware/rtx-3090). Trains Nerfacto on 500+ photos in 20-40 minutes. The 24 GB VRAM handles high-resolution scenes with complex view-dependent effects (reflections, refractions, transparency) that Gaussian Splatting struggles with. For research and VFX (where rendering quality matters more than real-time speed): NeRF on 24 GB is the gold standard. Total: ~$1,800-2,200. For production rendering: RTX 4090 ($2,000) renders novel views at 20-30 fps after training. NeRF is quality-over-speed; GS is speed-over-quality. Choose based on your use case.

The mistake: Training a NeRF on outdoor photos taken over 30 minutes, then getting a blurry, ghosted reconstruction. Why it fails: NeRF assumes a static scene. Outdoor lighting changes over 30 minutes (sun moves, clouds pass, shadows shift). Each photo has different illumination. The model tries to reconcile contradicting pixel colors — was that wall bright (direct sun) or dark (cloud shadow)? It averages them → blur. The fix: Capture photos within 2-5 minutes, ideally on overcast days (diffuse, consistent lighting). For sunny days: shoot within 1-2 minutes or accept artifacts in shadow regions. For dynamic scenes (people, cars), use robust NeRF variants that model transient objects (NeRF-W, RobustNeRF). Lighting consistency is the hidden requirement of photogrammetry. Same lighting across all photos = sharp reconstruction. Moving sun = blur.