Demo available
Learning what you like
Show the model a few examples you like and a few you do not.
blue mugs
red mugs
another blue mug
Show it what you like, even when the style is hard to describe. The model looks for patterns in your ratings. It can then suggest something new, sort existing choices, or adjust a design toward your taste.
A suggestion is still a guess. You decide whether it is any good.
Where this could be useful
Learn from ratings to rank choices, suggest something new, or make a small edit. These are possible applications; the current demo teaches the loop using drawings.
With a compatible generator, the small model can score or adjust its internal inputs before rendering. It can also rank things that already exist. These are the original generate, score and edit uses of the same learned preferences.
- Music and playlistsLearn what you like to hear next.
- Art and visual designChoose a direction by rating examples.
- Architecture and interiorsExplore layouts and rooms that fit your taste.
- Dating through mutual preferencesRate generated faces. Reveal real profiles only by mutual agreement.
- A feed without repeated promptingLearn from what people choose to see.
- Makeup and appearance previewsExplore a look through small visual changes.
More use cases: everyday choices, creative work, science and engineering
Research and engineering
- Molecule and drug researchRank research candidates against a stated goal.
- Material researchChoose candidates under several requirements.
- Genetics researchPrioritize candidates for an independent test.
- Protein researchScore candidates without treating scores as proof.
- Engineering with constraintsKeep useful designs within fixed requirements.
- Patient preferences in approved optionsRecord what matters to a patient.
Creative work
- Stories and interactive fictionLearn which style of story a reader enjoys.
- Game levels and difficultySuggest levels a player wants to play.
- 3D assets and objectsChoose shapes that fit a project.
- Video edits and effectsLearn which cuts and visual styles work.
- Voice and narration styleChoose a comfortable pace and delivery.
Design and everyday choices
- Fashion and outfit choicesFind clothes that fit your style.
- Automotive designExplore car shapes, colours and cabins.
- Interface and layout choicesLearn which layouts someone finds usable.
Personal choices
- News and content curationFind useful articles in a preferred format.
- Recipes and meal choicesLearn taste within the ingredients you choose.
- Learning materialsFind explanations a learner finds useful.
- Product recommendationsRank available products by personal taste.
Language and image generation
- Stop an image that misses your tasteSpend rendering time on more promising ideas.
About the original project
SkinDeep first tried this with generated faces in 2018–19. The current browser demo uses simple drawings so you can see what changes. Real image generators and human preference studies need separate tests.