Editing Neural Radiance Fields and 3D AI (section)

== <span style="color: #FFFFFF;">Applying</span> ==
'''3D Gaussian Splatting reconstruction:'''
<syntaxhighlight lang="python">
# 3DGS pipeline: capture images → COLMAP poses → train Gaussians → render
# Using the original 3DGS repository or simplified libraries

# Step 1: Prepare input images and estimate camera poses with COLMAP
import subprocess
def run_colmap(image_dir, output_dir):
    subprocess.run(["colmap", "automatic_reconstructor",
                    "--workspace_path", output_dir,
                    "--image_path", image_dir,
                    "--camera_model", "SIMPLE_RADIAL"])

# Step 2: Train 3D Gaussians (using gaussian-splatting library)
# from gaussian_splatting.train import train_gaussians
# gaussians = train_gaussians(
#     colmap_path="./colmap_output",
#     output_path="./trained_scene",
#     iterations=30000,
# )

# Using nerfstudio for a high-level interface
def train_nerf_scene(image_dir: str, method: str = "splatfacto"):
    """
    Train a NeRF or Gaussian Splat scene using nerfstudio.
    method: 'nerfacto' (NeRF), 'splatfacto' (Gaussian Splatting)
    """
    # ns-process-data images --data {image_dir} --output-dir data/processed
    # ns-train {method} --data data/processed
    # ns-render camera-path --load-config outputs/*/config.yml \
    #     --camera-path-filename camera_path.json --output-path render.mp4
    pass

# Instant-NGP training (much faster alternative)
# import pyngp as ngp
# testbed = ngp.Testbed()
# testbed.load_training_data("./transforms.json")  # COLMAP → transforms format
# testbed.train(1000)  # Trains in seconds!
# image = testbed.render(width=1920, height=1080)
</syntaxhighlight>

; 3D AI tools and frameworks
: '''NeRF training''' → Nerfstudio (nerfacto, splatfacto), Instant-NGP (NVIDIA)
: '''3D Gaussian Splatting''' → Original 3DGS, Gaussian Opacity Fields, Mip-Splatting
: '''Text-to-3D''' → Shap-E (OpenAI), Wonder3D, Zero123, DreamFusion
: '''Pose estimation''' → COLMAP (SfM), PixSFM, HLoc
: '''Real-time rendering''' → WebGL exports from Gaussian Splatting; NeRF→mesh conversion
</div>

<div style="background-color: #8B4500; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">