4 ms·
> Stable diffusion does not default to creating black people (you have to insert black into the prompt), honestly in order to make race an irrelevant or random
by pmarreck 4y ago
> Stable diffusion does not default to creating black people (you have to insert black into the prompt),
honestly in order to make race an irrelevant or random factor, I believe the training data would have to be evenly divided between all known races... and possibly some blends of races? And I think that would be a good thing.
Could you synthesize such data by (ironically perhaps) using AI to change the race of a person in a given training photo?
- AstralStorm 4y agoYou can certainly try this kind of data augmentation strategy. Pretty sure it'll fail because of the text analysis still being biased. You would need equalized amounts of text referring to various groups of people and topics too. That's much harder to augment.
- pmarreck 4y agoWhat's stopping someone from taking training text that mentions races by name (using any common name) and having it automatically changed to a different race for a new training sample?