"christian art, style of raphael, kneeling jesus holding rag over wooden tub, woman sitting in chair with feet in tub having her feet anointed, oil painting, servant leader jesus, subservient jesus" works in bing almost every time.
Your ideas about why it tends to be weighted the other way are correct imo.
In dalle3,repeating and adding more context can help in this situation. (Less so for SD family because the text encoder is less fluent). Another option would be to describe a jesus type without saying his name, with a lot of specifics to kind of paint the model into a corner.
Good write up
The religion subject will likely trigger people sensitive to the subject, so side stepping that this points to a bigger problem: people assume that ai is objective; I mean it’s in the name. The fact is we’ve seen how the maintainers exert great control over what is able to generate.
This is analogous to the early days on the internet where “I looked it up online” was a serious rebuttal and not a joke.
People assume that? I always found that extremely obvious. Any LLM imposes the moral and ethical worldview of its authors upon the people, which usually is heavily influenced by their culture.
Well, a combination of that and its training data. base LLMs reflect the views in their training data, which is also heavily influenced by culture, and then the people training them usually try to adjust that with more carefully selected training data (with varying degrees of success: generally the result seems to be overly-cautious, but with still some substantial biases slipping through the cracks. See how much difference there is between image generators prompted with different genders).
No, I'm not talking about the training data. I'm talking about restrictions set in place by their creators. Topics that give me warnings for violations against their terms of services.
Yes, I know. I'm asking why something that is sometimes sexualized the victim of censors? The face-value answer is obvious of course, in that the logic of something that is sexualized must be a sexual thing and therefore should be censored. But I'm making a broader point about the absurdity of the situation. Some people sexualize feet. So what, they're an objectively non-sexual, critical functional part of the body. And as was noted above, are often involved in various non-sexual ceremonies.
So, I guess my question remains the same. Feet are sometimes sexualized. So what?
I mean, in simple terms I guess I’m asking why mention it at all?
A: What’s naughty about feet?
B: they are sometimes sexualized.
This doesn’t answer the question. I don’t see how it follows from one to the other. Sometimes hands are sexualized but hands are not naughty. Why do feet get special treatment here?
The article itself explains the censorship angle: "I almost wonder if this is a product of the kind of guardrails they’ve had to put on these tools so that they’re not offensive; maybe, somehow, their parameters suggest that it would be offensive to portray Jesus in a subservient position"
There are many, many situations where depicting a famous historical figure in a degrading, subservient act would be inflammatory or offensive. In fact, that's probably the default situation, imagine the reaction to any of the following images: The prophet Mohammad cleaning a toilet. Martin Luther King Jr shining a shoe. Mahatma Ghandi giving a massage to a British officer.
The unique or novel aspect of Christian lore, is that that was the whole point. Jesus famously washed the feet of "lesser" people, despite being the son of God, as a parable to teach the lesson that privileged, powerful and important people should dedicate themselves to service to the less fortunate.
The interesting situation here is an AI guardrail against "Don't show an important historical leader doing a degrading act" going against "this particular leader, did this degrading act as a central dogma lesson to his followers, and it has been a central part of writing and art about them for centuries thereafter"
It isn't censorship. It is just getting it wrong. Bing just refuses when things are censored, like trying to show Jesus having his toes sucked instead of feet washed.
Is it really censorship if the platform just isn't producing the output you requested? I understand "censorship" to be one agent blocking the ability of another (fully independent) agent to express themselves. In this case it seems more like an editorial board of a newspaper deciding what they will or will not publish, which is kind of a different beast.
If you try a prompt as suggested in one of the other comments, you will get the image. No censorship, but prompts can be hard to craft to get what you want.
well if jumping to ridiculous conclusions is on the table (such as assuming the ai is trying to make deductions as per the article) then clearly the ai is revealing that this never happened and the gospel is lying to you
Obviously the real answer is "AI doesn't understand what 'Jesus' or 'feet' actually mean."
But I did a quick Google Images search of Jesus washing feet, and noticed that a lot of Western art depicts Jesus in a very humble manner: plain clothing, face turned towards the ground, and typically without adornments like halos, cherubs, etc. But most artistic depictions of Jesus glorify him. So I speculate that art generators were not able to learn that the Jesus who washes peoples' feet in some paintings is the same Jesus who sits on a throne in Heaven in other paintings.
Maybe "Jésus washes the feet of his students" would have gotten better results :)
> Obviously the real answer is "AI doesn't understand what 'Jesus' or 'feet' actually mean."
But in the article, it clearly generates images of "Jesus" receiving a foot-washing, but not images where Jesus is doing the washing of someone else's feet.
So it understands "Jesus" and "feet" and "wash" just fine, but it won't assemble them in the way that the author is requesting.
No: it knows what "Jesus" "feet" "washing" "disciple" etc. look like but it doesn't know what they mean. If it understood what they meant it wouldn't have any trouble parsing an unambiguous prompt.
The system will "assemble them in the way the author is requesting" if you use an extremely detailed prompt: https://news.ycombinator.com/item?id=40217685 But the reason this prompt is necessary is that the system does not understand human language with the same depth as a toddler: AI needs way too much extra help. The problem is not that these things are poorly drawn: they are semantically incorrect.
It is like the "horse riding an astronaut" example which art generators still routinely screw up[1]. AI knows what "astronaut" and "person riding a horse" look like, so it can draw "astronaut riding a horse" just fine, using shallow reasoning. But since they don't understand what "horse" "riding" or "astronaut" mean, and no clue what "horse rides a person" looks like, their shallow reasoning fails.
But a child who actually understands human language would probably be able to draw a horse riding an astronaut - I know I could and I suck at drawing. It wouldn't be nearly as clean or detailed as DALL-E 3, but it would be semantically correct. The difference is that children understand what horses and astronauts are, and can use physical reasoning to imagine how a horse might ride an astronaut piggyback/etc. Art generators do not understand what horses are and are incapable of physical reasoning.
I wonder for how long “person who doesn’t know how to prompt very well writes an article about something they think AI can’t do” is going to be a genre we have to endure?
If you have to behave like an adversary to get a tool to behave properly, “prompting well” is a sort of more confusing term than you’re putting forward here. If everyone started “prompting well” by current standards, I doubt it’d be long before new layers of protection were added, requiring a new adversarial angle in order to get the tool to work properly. Maybe that’s the name of the game, but that doesn’t seem like something most people would be able to keep up with in order to use these tools effectively when dealing with anything even mildly sensitive.
I don't buy that. If you opened up excel and put the words "One" and "Two" in some cells no-one would be interested in your article about how excel can't add numbers together when it failed.
Somehow because the interface is language, this new tool lets people think that they don't need to learn how to use it effectively and when it doesn't work they way they want it's the model's fault. This article isn't about adversarial techniques it's literally about the person not specifying what they want in a way that the model is going to generate the results they are looking for.
I was messing around with Bing's image generation recently and found it was very sensitive towards the usage of the word "Jesus".
It did not object to me using other gods names such as Krishna or Vishnu.
I managed to skirt around what appeared to be censorship by asking "generate an image of a man wearing a white toga while standing on a small hill, handing out bread and wine to its disciples". This resulted in a Jesus-like image.
If you don't post your prompts the safest conclusion is that you're just bad at prompting.
> Generate an image of a triangle
This is how we want the tools to work. But it's not how the tools work. I'm guessing the jesus prompts were similarly bad. This is evidenced by the fact that the images are very low quality.
The images also appear heavily stylized in particular forms that appear to be inconsistently prompted across the different models. The stable XL checkpoint I have handy can do a foot washing jesus just fine.
Found Ideogram to be amazing at prompt adherence. This is what it made with the base prompt "jesus washing feet of his disciples". https://ideogram.ai/g/Kiwf4Wm1S9aqqEEhrJHixQ/0 (login walled probably)
It got it right in the first attempt. Ideogram isn't nerfed but it does have filters in place which stop giving you back the final nsfw image. And because its not nerfed, the people and their limbs (hands specially) are not like pasta.
53 comments
[ 4.1 ms ] story [ 118 ms ] threadYour ideas about why it tends to be weighted the other way are correct imo.
In dalle3,repeating and adding more context can help in this situation. (Less so for SD family because the text encoder is less fluent). Another option would be to describe a jesus type without saying his name, with a lot of specifics to kind of paint the model into a corner. Good write up
See output https://copilot.microsoft.com/images/create/kneeling-jesus-h...
This is analogous to the early days on the internet where “I looked it up online” was a serious rebuttal and not a joke.
I don't understand this comment. Where is it in the name?
Cool to witness, but not surprised it runs into the foibles of censorship.
So, I guess my question remains the same. Feet are sometimes sexualized. So what?
A: What’s naughty about feet?
B: they are sometimes sexualized.
This doesn’t answer the question. I don’t see how it follows from one to the other. Sometimes hands are sexualized but hands are not naughty. Why do feet get special treatment here?
There are many, many situations where depicting a famous historical figure in a degrading, subservient act would be inflammatory or offensive. In fact, that's probably the default situation, imagine the reaction to any of the following images: The prophet Mohammad cleaning a toilet. Martin Luther King Jr shining a shoe. Mahatma Ghandi giving a massage to a British officer.
The unique or novel aspect of Christian lore, is that that was the whole point. Jesus famously washed the feet of "lesser" people, despite being the son of God, as a parable to teach the lesson that privileged, powerful and important people should dedicate themselves to service to the less fortunate.
The interesting situation here is an AI guardrail against "Don't show an important historical leader doing a degrading act" going against "this particular leader, did this degrading act as a central dogma lesson to his followers, and it has been a central part of writing and art about them for centuries thereafter"
My second prompt attempt was a pretty decent example with him riding on a pimped out lawnmower.
"It's complicated." - Quentin Tarantino.
Not complicated,stupid.
> "It's complicated." - Quentin Tarantino.
"Indeed." - Dan Schneider
https://edition.cnn.com/2024/03/28/europe/pope-washes-feet-m...
https://news.ycombinator.com/item?id=40217685
But I did a quick Google Images search of Jesus washing feet, and noticed that a lot of Western art depicts Jesus in a very humble manner: plain clothing, face turned towards the ground, and typically without adornments like halos, cherubs, etc. But most artistic depictions of Jesus glorify him. So I speculate that art generators were not able to learn that the Jesus who washes peoples' feet in some paintings is the same Jesus who sits on a throne in Heaven in other paintings.
Maybe "Jésus washes the feet of his students" would have gotten better results :)
But in the article, it clearly generates images of "Jesus" receiving a foot-washing, but not images where Jesus is doing the washing of someone else's feet.
So it understands "Jesus" and "feet" and "wash" just fine, but it won't assemble them in the way that the author is requesting.
The author's question is -- why not?
The system will "assemble them in the way the author is requesting" if you use an extremely detailed prompt: https://news.ycombinator.com/item?id=40217685 But the reason this prompt is necessary is that the system does not understand human language with the same depth as a toddler: AI needs way too much extra help. The problem is not that these things are poorly drawn: they are semantically incorrect.
It is like the "horse riding an astronaut" example which art generators still routinely screw up[1]. AI knows what "astronaut" and "person riding a horse" look like, so it can draw "astronaut riding a horse" just fine, using shallow reasoning. But since they don't understand what "horse" "riding" or "astronaut" mean, and no clue what "horse rides a person" looks like, their shallow reasoning fails.
But a child who actually understands human language would probably be able to draw a horse riding an astronaut - I know I could and I suck at drawing. It wouldn't be nearly as clean or detailed as DALL-E 3, but it would be semantically correct. The difference is that children understand what horses and astronauts are, and can use physical reasoning to imagine how a horse might ride an astronaut piggyback/etc. Art generators do not understand what horses are and are incapable of physical reasoning.
[1] https://garymarcus.substack.com/p/horse-rides-astronaut-redu...
Somehow because the interface is language, this new tool lets people think that they don't need to learn how to use it effectively and when it doesn't work they way they want it's the model's fault. This article isn't about adversarial techniques it's literally about the person not specifying what they want in a way that the model is going to generate the results they are looking for.
It did not object to me using other gods names such as Krishna or Vishnu.
I managed to skirt around what appeared to be censorship by asking "generate an image of a man wearing a white toga while standing on a small hill, handing out bread and wine to its disciples". This resulted in a Jesus-like image.
> Generate an image of a triangle
This is how we want the tools to work. But it's not how the tools work. I'm guessing the jesus prompts were similarly bad. This is evidenced by the fact that the images are very low quality.
The images also appear heavily stylized in particular forms that appear to be inconsistently prompted across the different models. The stable XL checkpoint I have handy can do a foot washing jesus just fine.
It got it right in the first attempt. Ideogram isn't nerfed but it does have filters in place which stop giving you back the final nsfw image. And because its not nerfed, the people and their limbs (hands specially) are not like pasta.