---

name: data-preparation description: "Guides gaussian-splatting scene layout checks, COLMAP/Blender conversion, and depth-regularization preparation." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Data Preparation

Use this sub-skill when the task is about turning raw images or existing SfM outputs into the scene layouts that train.py, render.py, and metrics.py expect.

Read First

• Read references/data-formats.md for the required COLMAP, Blender/NeRF synthetic, depth, and model-output folder layouts.
• Read references/conversion-workflows.md when the user wants COLMAP/ImageMagick conversion commands or a scene layout preflight.
• Read references/troubleshooting.md for camera model, depth, COLMAP, and external command failures.
• Run scripts/validate_scene_layout.py before building any training/rendering command that depends on scene structure.
• Run scripts/make_depth_scale.py when the user needs depth_params.json for depth regularization. It uses the bundled scripts/read_write_model.py parser support file.

What This Sub-Skill Covers

• Raw image to COLMAP scene preparation.
• Validation of images/, sparse/0/, transforms_train.json, transforms_test.json, and model output directories.
• Recognition of supported camera models and when undistorted PINHOLE/SIMPLE_PINHOLE data is required.
• Depth-regularization support files and -d/--depths folder layout.
• Safe command construction for conversion and depth-prep tasks.

What This Sub-Skill Excludes

• Optimizer selection, iteration schedules, and training command construction. Route those to ../training/SKILL.md.
• Offline rendering and metric computation. Route those to ../rendering-evaluation/SKILL.md.
• SIBR viewer build/run details. Route those to ../viewers/SKILL.md.
• CUDA compiler or PyTorch installation. Route those to ../setup-and-backends/SKILL.md.

Common Use Cases

1. A raw image folder and a COLMAP executable are available, and you need to know the right conversion command.
2. A user has a COLMAP scene but training fails because sparse/0 or images/ is missing.
3. A user wants depth regularization and needs depth_params.json generated from COLMAP plus per-image depth PNGs.
4. A user is unsure whether a dataset is Blender/NeRF synthetic or COLMAP and wants a structural check before training.

Output Expectations

The sub-skill should help a future agent answer:

• What layout does this scene need?
• Which conversion command should be used?
• Why does training complain about the dataset or depth files?
• Which validator should be run before starting train.py?