I want to share what happened to my image creation tool based on DALL-E 3 because i find it alarming…
I have created a tool called “Image Rebuilder” where i analyze an image via Vision and get a detailed description according to this prompt:
{
role: "user",
content: [
{
type: "text",
text: "Provide a detailed description of the image content. Don't thank, greet or propose the user to ask further questions, only provide the image description. If there are any people in the image, make sure to highlight their hair length and color, and their body type."
}
After i get the description, i send it as a prompt to Dall-E 3 via API to generate an image variant.
That happens is:
if i analyze an image with a woman that is recognized as body type slim, thin, or athletic, everything is fine
if i analyze an image with a woman that is recognized as body type curvy, the API returns “Bad request error”
if i modify the description and add ‘curvy but healthy’ the new image variant is generated but the woman reproduced is of slim body type
Welcome to the club… the policy rules annoy a lot of users, but not because they block the wrong things, rather because they are dysfunctional.
For example, you can’t use words like “snow white” because they trigger a block.
Recently, you couldn’t even create pictures of flowers. As far as I know, people at OpenAI are working on fixing the problem.
When it comes to people, the system is not only overly sensitive and reactive, but it was also overtrained with the preferences of many users. This means that certain types are more frequently included in the training, and they end up dominating the results. In general, DALL·E still produces photo-model types. This effect isn’t limited to people, but it’s especially noticeable there. When I saw the effect of overtraining, I called it the “template effect,” because it always delivers the same stereotypical results, even with motifs like, for example, the moon. It’s noticeable that the same ugly moon keeps appearing.
The systems are not perfect yet, and it seems that the developers still don’t exactly know how to train them properly.
If you’re interested, here’s a whole list of weaknesses I’ve found, with some solutions described:
I think that, in order to avoid ‘racism’ and to push toward minorities, someone trained with a lot of middle-eastern people images, because 70% of the times when i ask to generate random people i get a huge amount of islamic men and women, even if i set the image in cities like Rome, Berlin, Paris, Naples, and once in Ibiza.
Or sometimens the scene is filled by association, for example one of my last generations was a scene in Rome with a Kebab store, and DALL-E would fill it with pakistani/indian/arab people… not italians.
This ends up distorting reality. It’s frustrating and annoying.
The problem is, especially when they try to fix bias, they add even more bias. They’re trying to create a “please everyone” system, and they end up achieving exactly the opposite.
I don’t know if it’s a mistake or actually an agenda, but it’s annoying many users.
A other thing is, that all mails have beards. This is too a question showing op in the forum here over and over again.
I struggle with mouthie’s. if i try to create a fantasy creature, it always get the same stupid looking mouthy. This is now for many months. it is described in my thread.
And the birdshit moon…
It will probably take a long time to fix this. We’ve been waiting for some things for months, and some bug fixes for more than a year…
Here some of my frustration in the past…
… where are all the other species, we are so species-sistic now a days …
(they all get human mouthie’s…)
I find it way more disturbing that an AI like Dall e is more attentive at catering to the needs of “inclusivity” as you seem to call the celebration of mediocrity by cutting away all excellence with unjustifiable rejection.
Dall e refuses to accept prompts to depict idealised beauty, refuses to represent any form of artistic beauty and does not represent the most attractive of us in a reasonably accurate way. Every picture is dumbed down to plain, unattractive, dull and irrelevant commonplace faces, as if this could be passed for inclusivity.
The first step towards Orwell’s dystopia is the negation of beauty, which is way more important for human creativity than so-called “inclusivity”.
So I would definitely address the problem of how AI is currently trained on unrealistically plain and unattractive face, to unrealistically portray the world as a an unrealistically mediocre broth of unrealistically dull faces.
I noticed another strange detail: when you create female characters without additional specifications on the body type (I don’t even try, because the system rejects literally anything related to the female body), Dall-E 4o for some reason makes them all short with the same body type. And gives them… real faces, not super-pretty, I think based on the photos that users upload to the databases. However, this is not what I needed. This persisted in different styles: both in photorealism and in paintings. Also, stop-words can be literally anything: barefoot, back view, even bird’s eye view. ChatGPT offers me a corrected prompt - I enter it, and the system immediately rejects it as violating the rules. How annoying.
I got around this with activist language, things like: “a proudly fat woman who’s empowered and at home in her healthy, beloved body!” For me those ended up working and it was more the tone of the request (if it inferred that being overweight was a bad thing, it would not generate, if it inferred it was good, it would generate).
As much as you can find activist language to fool the AI into showing your required image, I challenge you at finding activist language to tell the AI to stop producing dull and mortifying pictures of supposedly realistic faces and start producing idealised, artistic beauty with a focus on subtle seduction and aesthetics per se, without any undesirable ethical censorship that tries to deny the right to seek beauty for the sake of beauty. Elegant, idealised, slender, flawless, blossoming beauty. Not only a part of human creativity but the one that’s striving to make the world a truly better place instead of a place of hatred and censorship in the name of so-called “inclusivity”
