search

dmarx / whats-in-a-name

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

textual inversion + PEZ = prompt-space personalization #7

Open dmarx opened 11 months ago

dmarx commented 11 months ago

in theory, would be model agnostic as long as using same text encoder.

dmarx commented 11 months ago

as a starting place, could even just probe the top CLIP similarity names compared with the input "identity" image.

would be nice if I had the "identity filter" constructed. really just needs to be a simple MLP is all.

dmarx commented 11 months ago

separately: I wonder how PEZ does inverting prompts for strong identities like this?