Cleaned up training data, upressed images, removed lots of buckets to improve training. It's now much more reliable and stable.
Keep in mind that the model zooms in quite a bit for high quality details, so add a location description to get a more zoomed out view.
New keywords to play around with:
- closeup
- closed eyes
- eye_contact (tagged for the 2 women to look at each other)
- bukakke
- dripping

