In OpenAI’s case it is less about safety and more about serving outputs from an API without causing huge PR backlash and without having to wait for _proper_ safety research to be completed to do so.
Besides that, “safety” is a very abstract term, and when it comes to a specific model like this the problems become more prosaic things like “embarrassment”.
That could be of people you’ve made deepfakes of, but it could be OpenAI embarrassed that their outputs look bad, or their users embarrassed that they got an NSFW output for an SFW prompt.
I think the censorship is also about copyright though; it barely knows any characters while dalle mini users are generating stuff like “Shrek on trial for murder” all day.
Indeed and while I have no access to DALLE2, I’d be willing to bet it happily outputs Minecraft-style imagery and IP, as business-daddy MS has assured them it won’t be an issue.
Well “Minecraft” isn’t a character so there’s definitely some of that in it. It knows the look of other IPs too, like various movies or games like Borderlands, it’s just that it doesn’t know many people’s names whether they’re fictional or not. So maybe “Minecraft Steve” wouldn’t work - can’t tell atm as I’m out of credits.
From earlier prompts I tried, “Minecraft” makes things look conceptually blockier rather than visually blockier, if that makes sense. So a “Minecraft mansion” will turn into a midcentury-modernist house with flat walls and roofs rather than being made of voxels.
That’s a pretty neat effect, but if you use names of video games in dalle or other image AIs they tend to start generating UI elements on top of your image and it looks bad.
This may have more to do with them filtering out training data containing names than anything else. They did the same thing for their released GLIDE checkpoint (predates DALLE2 by a couple of months).
