search

dmarx / whats-in-a-name

[WIP] probing identity and bias in text to image models
MIT License
1 stars 1 forks source link

seed-constrained identity strength measure #25

Open dmarx opened 10 months ago

dmarx commented 10 months ago

related: #1

super simple test here: prompt <name>, <modifier>, e.g. Annalise Basso, colorful hair. with a fixed seed, uninformative names will produce basically the same output image, informative names will produce more unique results. do this for a handful of seeds, the "lift" relative to some baseline uninformative prompt (or the mode of the "uninformative" distribution) is the measurable strength of the identity

dmarx commented 10 months ago

the more i play with this, the less confident i am that it will work the way i hope. might work better in tandem with openpose controlnet conditioning for facial portrait