# Image Generation Benchmark

This subfolder contains the canonical train/test split for the glaze image generation benchmark.

## Files

- `train/images/`
- `train/metadata.json`
- `train/recipes.json`
- `train/targets.json`
- `train/sample_ids.json`
- `test/images/`
- `test/metadata.json`
- `test/recipes.json`
- `test/targets.json`
- `test/sample_ids.json`
- `dataset_statistics.json`

## Current canonical counts

- Train total: 4,490
- Test total: 443
- Train transparency labels: 2,968
- Test transparency labels: 344
- Train surface labels: 3,381
- Test surface labels: 331
- Train color family labels: 4,483
- Test color family labels: 443

## Benchmark usage

This track supports conditional image generation from glaze-related conditioning signals, including:

- RGB color targets
- transparency labels when available
- surface labels when available
- recipe composition and ingredient lists
- firing condition metadata when available

## Notes

- The image generation benchmark is separate from the larger property prediction benchmark.
- `metadata.json` is the richest entry point for multimodal loading because it links images, attributes, and recipe-side information.
- Missing `transparency` or `surface` values should be treated as unlabeled rather than negative labels.
- A minimal retrieval-style reference is provided in `../baselines/image_generation_baseline.py`.