Wouldn’t something like isMale*P(male=.66) work fine?
I wonder how you would generalize this?
The issue is the command "draw a judge" shows a white guy 100% of the time.
if you automate prompt generation to be "draw a {gender} judge", with as you suggest the term gender has 66% chance of being "male", sure.
But how do you extend that automated prompt generation to include "draw a {skin color} judge" or "draw a {gender}, {skin color} judge".
or to occasionally generate a character from the Book of Judges.
Or Judge Dread.
It doesn't think like that.
If it did, they could've just done `P(hasFiveFingersPerHand)=0.99999`.
But it doesn't even necessarily draw what you ask it to. Instead, it generally adopts a set of de-noising transforms that it's been trained to believe would tend to lead to what the prompt sounds like.. then whatever those transforms produce would, hopefully, be sorta like what was requested.
https://colab.research.google.com/drive/1dlgggNa5Mz8sEAGU0wF...
You can see them define a custom color loss and apply it simultaneously with the regular diffusion loss. I've actually expanded this notebook to allow regional specification of the custom loss.
It's quite difficult to define a function that detects if an individual has 5 fingers or not. That's the real issue.
My point was that it doesn't actually think like that. For example, prompting StableDiffusion for a picture of a doctor doesn't necessarily get it to draw a human at all, much less a doctor of a pre-determined sex; instead, StableDiffusion de-noises the image until the result emerges, where that result would (ideally) contain a doctor of whatever sex it happened to come up with.
That said, you're right that we can add more code to try to guide things.
We could even just brute-force it by just re-generating images over-and-over, or tweaking them after generation, until they match exactly what we wanted. (Realistically, something like branch-and-bound would probably be preferred to blindly guess-and-check-ing.)
And I think if you used a perfectly balanced dataset for training, you’d get these guardrails for free because the right probabilities would be baked into the model’s weights.
For example, say someone wants to generate a "US President" -- what would the ideal range of outputs be?
The article checked for just two things: sex (male or female) and skin-tone (I, II, III, IV, V, or VI). To date, all US Presidents have been male, and they were probably mostly skin-tones I or II (not bothering to check), except for Obama who was probably.. like IV or something (still not bothering to check).
So if we run StableDiffusion for a "US President", what would a "perfectly balanced" output look like? Should there be any women? What about the skin-tone distribution?
Also, Obama was a 2-term President, so.. if his skin-tone should somehow affect the distribution, should it have a stronger effect because he was in office for longer than average? Or should all US Presidents have the same effect regardless of their time in office? And either way, why?