DALL-E 3 blows other image generators away when it comes to understanding complex prompts. Why did they have to cuck it so hard with censorship? It's absolutely absurd, it can literally only generate G-rated images. Surely they must realize they are ruining their own creation. Why is it so censored and will they ever roll back the censorship?
What loopholes do you guys use to get around the censorship?
once it gets good enough to actually replace humans they won't let you use it at all. enjoy it while it lasts.
Dalle3 as a model doesn't have censorship, maybe just bias depending on dataset they fed it. However, you don't interact with it directly, GPT4 rewrites your prompts with it's globohomo/safety/diversity instructions and only its output is being fed into D3, not yours.
>GPT4 rewrites your prompts with it's globohomo/safety/diversity
Is that why whenever I ask for a group of people and then click on the image, the prompt was rewritten to say "a large group of people of various ethnicities and genders"?
"Open"AI is understands that removing NSFW from training dataset will decrease its capabilities significantly, but creating prompt blacklist would require a lot of work and constant curation, as new workarounds would be found at much higher pace then currently. That's why you aren't even allowed near the model itself.
Except the fact they don't face any backlash. If you want NSFW, set up SD and go full hog with ControlNet and inpainting. Sure, it is much more convoluted as you basically tard-wrangle whatever flavor of NAI-leak finetune, but YOU have full control and authority.
>That's what "Open"AI should be about.
Haha, no. "Open"AI is about proving a concept and finding investors, sorta like early Internet before dotcom bubble. They are a black hole on Microsoft's finances, but due to potential and expertise, they can burn Bill's cash for some foreseeable future. Smaller models with curated datasets and very specific usecases will stick though, we just haven't reached this state yet.
This stuff is wild. I really hope it's just hallucinated nonsense and not really the best way Microsoft(R) OpenAI(R)'s world-class ML engineers have found to control what a fricking probabilistic model does.
This just in, no-one in ML knows dick all and they just guess/hope for the best and shoot huge datasets at the problem and hope for the best.
Nothing like a picture of fantasy adventurers visiting the palace to see the public figure and public figure.
>Why did they have to cuck it so hard with censorship?
Wrongthink image created by their generator makes morons on Twitter angry. morons on Twitter are most easily swayed and manipulated into being good consumers who swear allegiance to a brand that stands up against wrong think. Therefore they are on the side of the morons. Just like every other advertiser.
Dalle is higher now but SD has higher ceiling long term, Dalle can't be improved because there is an ideology behind it in a way SD does not have.
>Dalle can't be improved because there is an ideology behind it
This SHOULD change though. I'm hoping they'll face enough backlash that they'll lift the ridiculously strict guardrails and just let people create pretty much anything that isn't illegal. That's what "Open"AI should be about.
Nope. They don't give a shit about false positives.
They would have 99.999% false positive rate if it meant the filter had the target success rate against actual misuses. False positives mean nothing, only the ability to filter images they really don't want getting out. If it means you can only prompt forests and pictures of buildings, then so be it.
It really depends. There are prompts that can make Dall-E suddenly go. "But you shouldn't have clothes!", usually giving you a naked dude. Not these, however. Though Dall-e, unlike SD, isn't specifically trained on nipples, or so it would seem.
https://files.catbox.moe/gskljc.jpeg
https://files.catbox.moe/eevc9c.jpeg
https://files.catbox.moe/il2xvg.jpeg
https://files.catbox.moe/mcpdty.jpg
https://files.catbox.moe/7bekgu.jpg
https://files.catbox.moe/x9bqqs.jpeg
https://files.catbox.moe/xqyd1b.jpeg
Kinda horrifying ngl
https://files.catbox.moe/mllita.jpg
Got hit with "Uploading files from your IP range has been blocked due to abuse." after that post despite not having uploaded anything BOT.
What the kek? BOT is shit. I quit.
Idk, stable diffusion does an amazing job if you have half a brain.
its no better than stable diffusion fricking moron