Apart from that it produces insanely convincing results. I believe people would not bat an eye if they'd get added in an update to iOS!
Then I tried "day gecko" and the output absolutely rocked. So I guess my first try was unlucky.
And then, yes, the background removal takes away half of these and leaves smoke.
Quite a shit thing to be happening.
Rather than impugning the model makers for not curating societies content to erase its biases, to my mind it demonstrates what’s broken with society, and should be used as an indictment of how we encode our society in our media - the fact that if you asked an oracle a question it produces racist output is more an indictment of society itself because the media of our society IS racist. Sweeping it under the rug serves no one, IMO.
Instead the story is “AI models are racist,” which misses the real problem. The real problem is that when a human in our society wants to portray a robber, they use a black man. That should be the story, and pitching it as a flaw with AI models is like criticizing the color of the paint when the foundation is cracked.
https://techpolicy.press/researchers-find-stable-diffusion-a...
Saying their product perpetuates stereotypes but ignoring its reflecting the entirety of medias bias is ignoring that all media perpetuates these stereotypes, and their models are no more or less perpetuating of these stereotypes than literally the entirety of all media. The fact they hold it up for careful examination in an irrefutable way is a feature in my mind, not a flaw.
I would note that SD produces a base model, which can and is routinely fine tuned. I would rather see a fine tuning that eliminates the bias in the media than see a base model that is fundamentally divergent from the state of the media today. That’s the proper abstraction - a base model that’s the basic output from mass training on available media, and models that specialize for some curation. But I also object to the base models being censored. The reason why is it cuts off a base truth from the semantic models underlying such that it’s outputs are at odds with observable reality. Specialized models shouldn’t be doing things like un censoring, but adjusting base truth to curated views. “Unbiased model” should be a fine tuning of reality. “Safe for work model” should be a fine tuning of reality. The challenge is the model producers don’t trust the model users to be adults making adult decisions and thinking adult thoughts - including about biases, stereotypes, etc.
But regardless, I think base models should never be thought of as final products but as a basis to produce a final product.
Not being further up on a hierarchy of many bad actors doesn't absolve you from responsibility. What goes into these models is an editorial decision, as is which one to use in your software. This author isn't distributing models and they aren't being held to task over the model's content— they're generating images from those models and distributing them, and those images are what bother people. If they used the same model but somehow didn't get objectionable results, nobody would know, let alone care.
You say the media is to blame? Sure. When you start generating content with a particular perspective, you are media.
Snakes are, at times, quite dangerous. Granted, not all snakes, but in terms of being risk-averse, and unaware of which snakes are quite possibly deadly, am I not, at least in my current ignorant state, best served to avoid all snakes?
Feel free to substitute "snakes" with any existing stereotype. I personally am a fan of "people who eat pizza with pineapple".
It's not one way though it perpetuates the bias. I'm not blaming the author of this as its Stability AI that need to clean their data.
We’re on it – we’re reviewing the model and fixing the issue ASAP. Thanks for pointing it out and for your patience!