Live browser demo
Learn more than one favorite style
Compare example selection and simple preference models.
This small interactive model has no human-preference benchmark. The optional example ratings are synthetic.
You might like two different styles and dislike what lies between them. Rate the mugs to compare a straight decision boundary with a small model that can learn a curved one. Both learn from the same ratings, here in your browser.
What learns?
Both models fit logistic regression to your ratings. The straight model reads color and width. The curved model also reads their squares and product. These nonlinear features can represent separated liked regions; this is not a mixture of large image generators.
Random selection samples an unrated mug. Least-certain selection chooses the unrated mug closest to the selected model's 50% score. Neither strategy is guaranteed to learn faster from your choices. Scores are model predictions, not measured accuracy or calibrated probabilities.
The two-style example supplies 12 explicitly synthetic ratings, preferring the red and blue ends while disliking the middle colors. It is a controlled illustration, not human preference evidence. Use Start over to provide only your own ratings.