53 lines
1.7 KiB
Markdown
53 lines
1.7 KiB
Markdown
# Training Sets
|
|
|
|
Brain training reference images organized by domain. All images are CC0 / public domain.
|
|
|
|
## Directory Structure
|
|
|
|
```
|
|
training_sets/
|
|
├── sunsets/ ← SKY domain: sunset, golden hour, sky gradients
|
|
├── oceans/ ← SEA domain: ocean, waves, coastal
|
|
├── food/ ← FOOD domain: food photography, plating
|
|
├── portraits/ ← PORTRAIT domain: people, faces, characters
|
|
├── urban/ ← URBAN domain: city, architecture, streets
|
|
├── nature/ ← NATURE domain: mountains, forests, landscapes
|
|
└── weather/ ← WEATHER domain: fog, rain, snow, storms
|
|
```
|
|
|
|
## Usage
|
|
|
|
From the `ralpha` repo:
|
|
|
|
```bash
|
|
# Run batch training on a domain
|
|
python -m brain.batch --dir /path/to/ralpha-ue5/training_sets/sunsets/ \
|
|
--domain SKY --llm gemini --max-iterations 20 --shuffle
|
|
|
|
# Or symlink for convenience
|
|
ln -s /path/to/ralpha-ue5/training_sets ~/ralpha/training_sets
|
|
```
|
|
|
|
## Sourcing Images
|
|
|
|
```bash
|
|
# Download CC0 sunset images (PxHere, with EXIF)
|
|
cd /path/to/ralpha
|
|
python scripts/dev/download_training_sunsets.py --output /path/to/ralpha-ue5/training_sets/sunsets/
|
|
```
|
|
|
|
## EXIF Data
|
|
|
|
Images with intact EXIF are preferred — the brain extracts:
|
|
- **Camera model + lens** → locks CineCamera focal length, aperture
|
|
- **Date/time** → computes exact sun position via ephemeris
|
|
- **GPS** → loads Cesium tiles for that location, locks sun azimuth
|
|
- **ISO/aperture/shutter** → exposure starting point
|
|
|
|
Images without EXIF still work — the brain uses VLM analysis to estimate scene parameters.
|
|
|
|
## Licensing
|
|
|
|
All images in this directory must be CC0 (Creative Commons Zero) or public domain.
|
|
The `manifest.json` in each subdirectory records the source and license of each image.
|