search

dmarx / whats-in-a-name

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

Try DIV metric from dreambooth #9

Closed dmarx closed 11 months ago

dmarx commented 11 months ago

rather than computing similarity to class centroid, they computed mean intra-class similarity for their diversity measure. they also define a "CLIP-I" measure for comparing similarity to real images, but this makes more sense in the personalization context I think than in a context where it's non-trivial to figure out what the relevant images in the training data were. I guess we could maybe use CLIP-retrieval to find those?

dmarx commented 11 months ago

i mean, yeah.. it's sort of cheating but fuck it, it works. https://rom1504.github.io/clip-retrieval/?back=https%3A%2F%2Fknn.laion.ai&index=laion5B-H-14&useMclip=false&query=bill+murray