SkinDeepRESEARCHSteve Seguin

Use case · Idea to test

Stop an image that misses your taste

Spend rendering time on more promising ideas.

  1. Learn from your image ratings
  2. Score a generation in progress
  3. Continue, cancel or restart
Illustration of the proposed workflow.

If a person consistently rejects a certain visual style, a generator could stop a poor candidate before finishing it and spend the remaining budget on another idea. This connects the original personal preference model to adaptive computation.

Where the small model fits

A personal scorer would learn how intermediate features relate to the person's final ratings. A separate stopping rule would estimate whether an initially poor candidate is likely to recover. A scorer trained only on finished images may not work on noisy previews.

What would need to work

Compare finished images people prefer per fixed time or compute budget, counting cancelled images and lost variety. Evaluate rejected candidates too, so good images cancelled early are visible. This is an untested extension of preference learning, not the current drawing demo.

Related experiments and evidence