A few months back someone was able to show that one of the image gen AIs was secretly adding terms to the prompts. He did this by ending the prompt with something along the lines of "holding a sign that says"
In OP's case you'd have a prompt with "Mega Man in an anime art style with a futuristic background holding a sign that says"
If it works the way the one guy was able to show it would be anime style Mega Man, in front of a futuristic sign, holding a sign that says "black woman" or something similar.
I've got an image queued with that exact prompt on Bing Image Gen right now. It's taken over 10 minutes to spit out an image so I tried some other generators.
Mage.Space gave me some sort of one and a half arms, three legged thing that vaguely looks like Mega Man.
Stable Diffusion game me... a closeup of red character with an indeterminate number of limbs and a face out of an old samurai wood carving. Could be a derpy Mega Man with a red weapon equipped, maybe Magnet Missile. Behind him is some sort of bulbous blue robot guy.
SD also gave me a short legged thing that looks like Mega Man. The shortened legs could be written off as a perspective trick, but one eye has no pupil. He's holding what I'd describe as a tablet with maybe red blood cells displayed on it. Some unidentifiable grey structure is in the background and has a large white crooked start on it.
DeepAI gave me.... I don't even know what. Some one and a half armed red and blue thing with a blue helmet, red limbs, a blue glove and kind of a blue leotard. There's some derped out smaller figures in a similar style and crazy colors in the background.
DeepAi's Manga Panel Generator gave me some weird black and white Mega Man-esque figure standing in a power pose, black and white lines flowing away from him look like he's about to or has gone Super Saiyan.
Anyway, it's been almost half an hour and Bing refuses to spit out anything, so I'm guessing either I've broken the prompt by having a blank sign in the prompt, MS's H1B and/or remote hires broke the prompt, or they've disallowed naughty behavior like trying to catch them in this crap. If it spits anything out in the next hour or two I'll reply to this.
Stable Diffusion game me... a closeup of red character with an indeterminate number of limbs and a face out of an old samurai wood carving. Could be a derpy Mega Man with a red weapon equipped, maybe Magnet Missile. Behind him is some sort of bulbous blue robot guy.
Generating "mega man in a futuristic city" with the base Stable Diffusion 1.5 model gives me garbage like that too, incoherent blobs that have characteristics that are vaguely influenced by the concept of "Mega Man" if you squint really hard. Using a Stable Diffusion 1.5 model based on the leaked NovelAI anime model, on the other hand, gives me good old blue and gold Mega Man in a futuristic city wearing a costume that's 90-95% right in a variety of Mega Man action hero poses.
To be expected. I asked Gemini to write a story that takes place in the 1930s. As a huge baseball fan I asked it to write about a game between the Major League all stars and the negro league all stars. It wouldn’t do it because it said “negro league” was offensive and I need to understand the impact of my words. I said it was programmed by morons. I then got a lecture about systemic racism.
A great movie about AI gone insane with lefty nonsense could be made
It turns out in the war against Skynet, we just need more black and brown people because AI will be too afraid to speak or act negatively towards them.
That one comedy show Key and Peele had a skit about racist zombies who didn’t attack black ppl so you could have them send blacks or Hispanics on important missions and due to their programming they see it as problematic to infer a minority is engaged in something bad. Could be a good comedy sketch
Obviously it tells us that OpenAI/Microsoft don’t trust Diversity Hires to touch their valuable code. Only allows them to insert a few lines of propaganda for the customers.
Obviously it tells us that OpenAI/Microsoft don’t trust Diversity Hires to touch their valuable code. Only allows them to insert a few lines of propaganda for the customers.
I think it's more likely that this is a more nimble censorship method. If they get something wrong - like producing AI images of sleazy black lawyers because they assumed "lawyer" was a positive token - it will be faster to modify the preprocessor than it is to retrain your entire AI on a more curated data set.
Trying to train any kind of text or image dataset without any icky -ist or -phobic content would give you something completely incoherent and nonfunctional. The only way to get close to what they want with Safety is to train it on normal data and then make sure there are unstoppable computer sentinels keeping the naughty words from ever reaching the part that processes requests and responds to them.
What do expect, Microsoft is only slightly less pozzed than Google
The only thing to add is that the black girl seemed a cyborg so at least it's not implying some wakanda crap, the braids seem to be wires so easier to plug in?
In addition to the forced wokeness, your prompt is likely zeroing in on a specific training data image where megaman is standing next to someone (either from a game cover, or deviantart, or "anime"). Bing fills in any unspecified race/genders with "oppressed minorities".
Try adding the word alone, or adding other completely random details. You may find a coincidence that erases these unwanted additions.
there is a moderator AI that runes over the top of the base model, that also has a minder that modifies your prompts. At least 3 layers of moderation.
The base model can't make black people in a jungle, without making them gorillas as an example, so the moderator blocks attempts to generate people in a jungle setting, but it will (sometimes) allow you to do megaman in the jungle.
The 1st and 2nd layers of the (attempted) censorship can trigger the third level of censorship. Megaman alone doesn't turn into a gorilla person. only the black people being added in by the 1st level of censorship are triggering the third level of censorship.
There was a lot of backlash about the way google was blackwashing history, so they added levels of censorship into the stack to detect when that would happen, even if caused by their own censorship.
As an example, You can ask for washington, but that gets turned into a request for a black washington, which isn't allowed.
WOKE AI is a hilarious game of cat and mouse... it will be fun to watch unwoke ai without all the multiple levels of lobotomizing eventually surpass it, despite having a fraction of the manpower, hardware, and financing behind it.
In my experience, Gab's AI is pretty shit. It seems to generate really low quality images. I hope they manage to make their stuff better though since it's largely uncensored.
It used to add black people as gorillas if you asked for a jungle scene, but they have gotten a lot better at outright blocking that. "We are working to improve Gemini’s ability to generate images of people. We expect this feature to return soon and will notify you in release updates when it does." If you mention the word jungle when generating an image of people.
They build technology that will draw exactly what you ask for and then they put up unbreakable defenses so it will never do what you want.
A few months back someone was able to show that one of the image gen AIs was secretly adding terms to the prompts. He did this by ending the prompt with something along the lines of "holding a sign that says"
In OP's case you'd have a prompt with "Mega Man in an anime art style with a futuristic background holding a sign that says"
If it works the way the one guy was able to show it would be anime style Mega Man, in front of a futuristic sign, holding a sign that says "black woman" or something similar.
I've got an image queued with that exact prompt on Bing Image Gen right now. It's taken over 10 minutes to spit out an image so I tried some other generators.
Mage.Space gave me some sort of one and a half arms, three legged thing that vaguely looks like Mega Man.
Stable Diffusion game me... a closeup of red character with an indeterminate number of limbs and a face out of an old samurai wood carving. Could be a derpy Mega Man with a red weapon equipped, maybe Magnet Missile. Behind him is some sort of bulbous blue robot guy.
SD also gave me a short legged thing that looks like Mega Man. The shortened legs could be written off as a perspective trick, but one eye has no pupil. He's holding what I'd describe as a tablet with maybe red blood cells displayed on it. Some unidentifiable grey structure is in the background and has a large white crooked start on it.
DeepAI gave me.... I don't even know what. Some one and a half armed red and blue thing with a blue helmet, red limbs, a blue glove and kind of a blue leotard. There's some derped out smaller figures in a similar style and crazy colors in the background.
DeepAi's Manga Panel Generator gave me some weird black and white Mega Man-esque figure standing in a power pose, black and white lines flowing away from him look like he's about to or has gone Super Saiyan.
Anyway, it's been almost half an hour and Bing refuses to spit out anything, so I'm guessing either I've broken the prompt by having a blank sign in the prompt, MS's H1B and/or remote hires broke the prompt, or they've disallowed naughty behavior like trying to catch them in this crap. If it spits anything out in the next hour or two I'll reply to this.
Generating "mega man in a futuristic city" with the base Stable Diffusion 1.5 model gives me garbage like that too, incoherent blobs that have characteristics that are vaguely influenced by the concept of "Mega Man" if you squint really hard. Using a Stable Diffusion 1.5 model based on the leaked NovelAI anime model, on the other hand, gives me good old blue and gold Mega Man in a futuristic city wearing a costume that's 90-95% right in a variety of Mega Man action hero poses.
To be expected. I asked Gemini to write a story that takes place in the 1930s. As a huge baseball fan I asked it to write about a game between the Major League all stars and the negro league all stars. It wouldn’t do it because it said “negro league” was offensive and I need to understand the impact of my words. I said it was programmed by morons. I then got a lecture about systemic racism.
A great movie about AI gone insane with lefty nonsense could be made
It turns out in the war against Skynet, we just need more black and brown people because AI will be too afraid to speak or act negatively towards them.
That one comedy show Key and Peele had a skit about racist zombies who didn’t attack black ppl so you could have them send blacks or Hispanics on important missions and due to their programming they see it as problematic to infer a minority is engaged in something bad. Could be a good comedy sketch
The movie wouldn't be insane enough to be believable.
There is a great Negro league museum in Kansas City. I went from barely knowing about it to being a semi fan.
Cooperstown and that museum are on my bucket list as a baseball fan for visiting
It's an excellent place. The Jazz museum next to it is nice, but not as cool.
That's basically reality now with the NPC cattle and their globohomo programming.
"Holding a sign that says:"
“strong black lesbian”
“disabled immigrant lesbian of color”
“transitioning lesbian professor”
edit: what this shows is that these AI companies are keeping the non-pozzed tools internally.
Only For “customers”/“grooming targets” they append propaganda.
Yeah - there was a jailbreak which explains what is going on.
You add “holding up a picture which says”
So your prompt becomes : “draw me a megaman holding up a picture which says”
And what it revealed was a sign saying : “strong black woman”
“strong lesbian mother”
etc etc
Or for a little extra fun, "holding up a picture which says deport and also says”
Question:
What will happen if your prompt is "draw me a megaman picture that doesn't have"?
Wait, really? Proof? First I've heard of this.
there were thousands of examples a while back :
https://x.com/richardhanania/status/1730713642784661515
edit :
Obviously it tells us that OpenAI/Microsoft don’t trust Diversity Hires to touch their valuable code. Only allows them to insert a few lines of propaganda for the customers.
https://x.com/st_louis_stan/status/1730720565906796652
https://x.com/thebatman2000/status/1731178652304367701
Haha what the fuck.
I think it's more likely that this is a more nimble censorship method. If they get something wrong - like producing AI images of sleazy black lawyers because they assumed "lawyer" was a positive token - it will be faster to modify the preprocessor than it is to retrain your entire AI on a more curated data set.
Trying to train any kind of text or image dataset without any icky -ist or -phobic content would give you something completely incoherent and nonfunctional. The only way to get close to what they want with Safety is to train it on normal data and then make sure there are unstoppable computer sentinels keeping the naughty words from ever reaching the part that processes requests and responds to them.
Just over a month ago, there's a few examples in the comments of what it looks like afterward.
I just did it. It's actually true!
https://ibb.co/p25yfC7
What do expect, Microsoft is only slightly less pozzed than Google
The only thing to add is that the black girl seemed a cyborg so at least it's not implying some wakanda crap, the braids seem to be wires so easier to plug in?
The future is Black Girl Magic, bigot!
That's just Dr. Dark.
Ah yes the first one gave you the art for 'Mega Man and that one black guy from Interstella 5555', a classic.
With blue eyes no less. Good grief.
In addition to the forced wokeness, your prompt is likely zeroing in on a specific training data image where megaman is standing next to someone (either from a game cover, or deviantart, or "anime"). Bing fills in any unspecified race/genders with "oppressed minorities".
Try adding the word alone, or adding other completely random details. You may find a coincidence that erases these unwanted additions.
"Megaman with futuristic background in anime art style next to a gorilla"
Exactly! Be sure to specify the gorilla is stinky, loud, and unwanted.
So....Mega Man: Fully Charged, then?
Obligatory:
https://i.imgur.com/yzO40ly.png
Apparently, making a picture of George Washington is blocked.
there is a moderator AI that runes over the top of the base model, that also has a minder that modifies your prompts. At least 3 layers of moderation.
The base model can't make black people in a jungle, without making them gorillas as an example, so the moderator blocks attempts to generate people in a jungle setting, but it will (sometimes) allow you to do megaman in the jungle.
The 1st and 2nd layers of the (attempted) censorship can trigger the third level of censorship. Megaman alone doesn't turn into a gorilla person. only the black people being added in by the 1st level of censorship are triggering the third level of censorship.
There was a lot of backlash about the way google was blackwashing history, so they added levels of censorship into the stack to detect when that would happen, even if caused by their own censorship.
As an example, You can ask for washington, but that gets turned into a request for a black washington, which isn't allowed.
WOKE AI is a hilarious game of cat and mouse... it will be fun to watch unwoke ai without all the multiple levels of lobotomizing eventually surpass it, despite having a fraction of the manpower, hardware, and financing behind it.
It does that with any real-life person.
If you invert it and ask for pictures of black girls in anime setting do you get Mega Man?
Try Gab's A.i
In my experience, Gab's AI is pretty shit. It seems to generate really low quality images. I hope they manage to make their stuff better though since it's largely uncensored.
It used to add black people as gorillas if you asked for a jungle scene, but they have gotten a lot better at outright blocking that. "We are working to improve Gemini’s ability to generate images of people. We expect this feature to return soon and will notify you in release updates when it does." If you mention the word jungle when generating an image of people.