Image Variation #3
Description
For images like the one we used in #2, the model isn't very accurate, and we end up needing to use more passes. My hypothesis is that this is because non-photographic images have a dramatically different Q-to-score curve than photographic ones, which we trained on with the gb82 dataset. To deal with this, we can:
- Integrate photodetect2, train a non photo model, and switch between them based on whether the source is photographic or not
- Train a smarter model with more terms that can modify its predictions based on image characteristics