Slider controlling the style direction between photo and illustrations.
Dataset is composed of 100 images, 50% photo, 50% illustrations.
Usage:
- Positive: steer the model towards photography.
- Negative: steer the model towards illustration.
"Limitations":
Due to the way it is trained (predicting the same image at the same noise and timestamp with 3 different captions) it will change the image a bit.
The "illustration" style might looks sloppy since the model might associate "illustration" with RL slop.
Placeholders replaced during training in captions:
[slider]
caption_placeholder = "<PHOTO_SLIDER>"
positive = "a real life photograph shot with an high high smartphone"
negative = "illustrated artwork, the illustration art"
neutral = ""