AI real-time human full-body photo generator
generated.photos
generated.photos
(And if you're wondering how it can look so good when 'everyone knows GANs don't work because they're too unstable', a widespread myth, repeated by many DL researchers who ought to know better, GANs can scale to high-quality realistic images on billion-image scale datasets, and become more, not less, stable with scale, like many things in deep reinforcement learning. See for example BigGAN on JFT-300m https://arxiv.org/pdf/1809.11096.pdf#page=8&org=deepmind , GigaGAN https://arxiv.org/abs/2303.05511 , Projected GAN https://arxiv.org/abs/2111.01007 , StyleGAN-XL https://arxiv.org/abs/2202.00273 , or Tensorfork's chaos runs https://gwern.net/gan#tensorfork-chaos-runs . 'Computing is a pop culture'...)
Are they using me to refine the model in some way?
Is that a wide spread myth? I thought it’s widely accepted that GAN is really good generating these artificial pictures (it’s what started DeepFake after all) when you know your model’s “button”. Similar to how this uses GAN since they have a model “boundary condition”. While humans are diverse, we have a set of repeatable features (two legs, two arms, etc). Diffusion models are great because you can control the latent space with something way more generic, like text hence why it’s been so much more mainstream.
Edit: actually I might be misremembering, I think Deepfake used VAEs?
You come across as severely, _severely_ biased and reactionary.
Get off of Twitter, maybe? I don't know but I have seen a ton of work done on large scale GANs. The incentives are pretty fucking obvious - they are faster to use. Of _course_ people want them to work for the same usecases as diffusion.
Maybe they will someday! But they aren't currently. You're referencing a zeitgeist that I am not familiar with and therefore come across overly defensive.
GigaGAN is the best quality out of those and requires 7 loss functions and is incredibly complicated.
Sure GANs can scale, but Diffusion models are drastically easier to scale.
Diffusion models are only 'drastically easier to scale' because researchers have spent the past 3 years researching pretty much nothing but diffusion models, discovering all sorts of subtle issues with them and how to make them scale, which is why it took them so long to become SOTA, and why massive architectural sweeps like https://arxiv.org/abs/2206.00364#nvidia were necessary to discover what makes them 'easier to scale'. If this level of brute force and moon math is 'easy', lord save us from any architecture which is 'hard'!
Reading your comments make me feel like that you believe that just every researchers (even extremely smart dude like Karras) just switch to diffusion models because they are idiots, they should have instead focus on GANs and today we will have GANs that are as powerful or more than the diffusion models we have today and also work one step; this is just a weird delusion. Diffusion models are just simply much easier to train (just a L1/L2 loss in most cases), write (for example your buggy BigGAN implementation), they usually work out-of-box on different resolutions and aspect ratios, you can just finetune them if you want to create an inpainting model; and for what is right now you just need much less compute to reach a good image coherency or maybe just reaching a coherence that as not been achieved by GAN models; like I would be curious even on a small scale experiment what a GAN (with ~55M parameters) would be able to perform after a 1-day/2-day GPU time of training on Icon645 dataset, because my diffusion model I can assure is much better than I could have imagine while being trivial to implement (I just implemented a Unet as I remember one, nothing rigorous and of course no architecture sweep).
> for example your buggy BigGAN implementation
The bug was that the batchnorm gamma was initialized to zero instead of one, so all the model params were being multiplied by zero instead of one. Literally any model that uses batchnorm is susceptible to this bug, and it certainly has nothing to do with GANs.
I think it’s too easy to spout dogma, and calling gwern delusional is something I’ve learned from experience will prove foolish as t approaches infinity.
Also no I don't take "he is gwern" as an argument ahah
This is what tends to happen when you find a superior method.
GAN’s are fine, they have plenty of promise for tasks requiring rapid inference. But diffusion models beat GANs on robustness and image quality every time.
I assume its cheap because they are burning money to build a business, its not fast at all, and the quality... sucks.
> And if you're wondering how it can look so good
I’m not.
I'm wondering why they’re trying to get people to use something worse than using a decent photorealistic SD1.5-based checkpoint with some basic prompt templating.
Not saying GANs can't be awesome, just that this site isn't what I’d use to make that case.
https://images.generated.photos/0wV1dBnZ15hGneEfqfZT7SdEIill...
My prompt was simply "Standing in front of a rocket.".
That looks more like a machine, an android perhaps, than an actual human.
It's not nightmare fuel it's a nightmare engine
So I was relatively unimpressed with the first few images, they looked incredibly unrealistic. But then after reading this message i decided to add "topless" to the notes just to see how crazy the result was.
I was actually impressed, it returned an image that actually looks lifelike once I gave it an adult request. Which reinforces the idea that this was optimized for porn.
For those curious, this was the result I was provided. NSFW of course, topless female gender (ai generated): https://generated.photos/human-generator/64e6c26e190809000fb...
[NSFW] https://generated.photos/human-generator/64e7e1aee4125d000cd...
And then I requested a full bush...
[NSFW] https://generated.photos/human-generator/64e7e47ff56326000b9...
>See for example BigGAN
I remember when you try training a BigGAN model on anime images, the quality was bad. Now look at this example, one single GPU, 1.5M images with a diffusion model: https://medium.com/@enryu9000/anifusion-diffusion-models-for... ,the difference in quality is absurd, you can say this or that is not true but the quality speak for itself, obtaining good quality on complex distribution is much easier with a diffusion model than a GAN.
For example in the case of the site linked they have conditioned the model on poses because you're not going to get anything close to be coherent without them with a simple StyleGAN as they say they're using.
Because there was a bug in the code, in a part unrelated to the GAN itself.
> the difference in quality is absurd
Yes, it does help to train on anime with code that isn't buggy. (BTW, Skylion was getting good results with GANs on anime similarly restricted to centered figures like those samples, he just refuses to ever publish anything.)
Also the bug is probably related to the added complexity of training a GAN model in comparison to a diffusion model.
I don't think anyone is wondering this, especially if they are used to playing with diffusion models.
"And if you're wondering how it can look so good when 'everyone knows GANs don't work because they're too unstable', a widespread myth, repeated by many DL researchers who ought to know better, GANs can scale to high-quality realistic images on billion-image scale datasets, and become more, not less, stable with scale, like many things in deep reinforcement learning."
My interpretation from how this is written is that you are saying: "Researchers say that GANs are unstable at inference time." Is that the correct reading? If so, where have you seen this sentiment? More commonly I've heard that people criticize GANs at inference time for mode collapse and monotony (which comes from training), not instability.
Or are you saying that researchers say GANs are unstable during training, which is a common criticism. Don't you feel this is the case? A lot of different tricks are used for getting generators that are roughly balanced with more powerful discriminators so that the adversarial game is balanced, like TTUR and encodec's weight balancer etc. In this case, are you saying that GAN training is as straightforward as diffusion training?
In my experience, GAN training involves an unholy number of heuristics and best-practices are still quite murky.
I am eager to hear your response.
[edit: my experience with GANs and diffusion models isn't about taking image problems with existing working models and scaling or refining them, but about applying GANs and diffusion models to different domains (audio) and on novel problems with different kinds of conditioning. I would love to see more controlled experiments comparing weak and sophisticated generative backbones as the experimental control and varying the training regime across different GAN + diffusion flavors.]
eg (NSFW!): https://generated.photos/human-generator/64e741e6c4c0f80009e...
Wrong number of limbs, wrong number of fingers on hands, etc.
The result is ok sometimes, but that's probably only about 5% of the time. Here's a decent result (very NSFW) without immediately noticeable deformities:
https://generated.photos/human-generator/64e7590e482612000b1...
Probably 95% of the images generated so far have been clearly deformed though.
We trained for 36 days on a v4-8 on 558k images.
https://nyx-ai.github.io/stylegan2-flax-tpu/
We were hopeful GANs would beat diffusion models when trained on specific domains. But we’ve now switched to Stable Diffusion and Dreambooth training which has proven more much efficient for this purpose.
I still have hopes for GANs! I miss their insane inference speed.
My only hope is that this extreme enshittification of online images will make people completely lose trust in anything they see online, to the point where we actually start spending time outside again.
[1] (NSFW, seriously.) https://deepnude.cc/
[2] (NSFW, seriously.) https://undress.app
[3] (NSFW, seriously.) https://porn.ai
This is just disgusting. I thought it would just be a uncensored generative AI but certainly wasn't expecting peeping-tom as a service. And advertise it so blatantly as being able to virtually strip any girl you have pictures of just makes me sick.
And the same would apply to doing this the old fashioned way in Photoshop, however you have to admit that taking it from "need special software and experience in using that software" to "just upload their image to a website and get back their AI generated nudes" is a huge change in how accessible this is.
If nude pics can get you fired, work culture needs to change.
Same with relationships.
Basically, people need to learn not to trust digital media at all without some kind of authentication, and to be a little more tolerant of nude human bodies when they do pop up.
Only because this tech isn't yet well known. People will just correctly learn to not trust anything they see online.
It's only Voice cloning that really scares me.
That's pretty much the behaviour of a scam website.
Well in a weird way it will provide cover. You could post nudes of yourself online and just explain it away as bad actors using AI.
A few years ago generating images like many of these was unimaginable, and now this does it, but perhaps 50% have some silly flaw in the image. Unless there were some reason to believe that further improvement isn't possible, as you say, it is likely to become indistinguishable from reality in the relatively near future.
In some ways, I think that actually makes it safer now… the more trivial it is the more people will stop trusting photos as automatically being real.
This will, if not already, make its way into porn and increase the amount of content there. It's already wild and with this, I don't expect anything good to happen.
I'm in full-bore accelerationist agreement with this point. Defense lawyers must love this.
Not really, depending on who you are "convincing". Up until relatively recently it was usually quite easy for photo experts (and often even less than experts, just people trained to look for certain "tells") to detect digital manipulation. But, on the flip side, yes, digital manipulation has occurred in the past, and I think it's a mistake to discount the strong negative effects it has had, e.g. many people having a completely unrealistic view of what most real humans actually look like (e.g. https://scottkelby.com/faith-hill-redbook-magazine-retouch-f...).
> The new technology just makes it cheaper for mischievous people to make fakes
That's a huge deal. Just look at the concerns around the use of LLMs to generate (and run) misinformation campaigns. Obviously those campaigns can and are run now, but the thought of it being incredibly cheap to do so, by people of extremely little skill, changes the information landscape drastically. Doing it for images is just another piece of the puzzle.
I don't think there is a nefarious purpose going on, i.e. getting people to sign up and stealing their info or payments, etc. However, it contributes to the erosion of trust on the internet. You're no longer sure if you're talking to a real dog in pajamas online or an AI pretending to be one.
Here's Kurt Cobain in a universe where he gets a regular office job, goes to gym, and never starts a grunge band...
https://generated.photos/human-generator/64e708a6190809000fb...
In Germany for example every commercial website (which is defined very broadly and applies to most websites) is by law required to have an imprint listing the person/org responsible for the website and how to contact them (an e-mail address is not good enoguh). This means you can go to any German business website and get the full address, EU VAT ID and registration number of that company.
Under the GDPR more broadly (which also affects foreign companies offering services to EU residents) every company is required by law to have a privacy policy and that policy must include whom to contact for concerns and requests regarding personal data of the user/visitor and who (which legal entity) processes and stores the data.
This is the opposite of doxxing. It protects private individuals by making transparent to them who they are interfacing with and who holds their data. This is necessary for informed consent.
Sidenote: the website/app's cookie notice is pointless as it's using the same "redirect to google.com if user says no" logic porn sites used to do (do they still do this?) for age checks. The app also works without accepting the terms, so either it can work without accepting them or (more likely) it doesn't actually wait for the user's consent. Either way it doesn't do anything and doesn't comply with any privacy laws I'm aware of that would require it.
It's the same attack, to deanonymize, to hunt people down from the internet because you don't like what they say or do.
Germany is the only country in the world with that Impresseum policy because of it's highly leagalistic Prussian background, and you would find that many in the hacking community (e.g. CCC) take huge issue with it
The CCC has a strong anarchist tendency unlike the US tech bubble which has a more libertarian (i.e. free enterprise) streak. They absolutely do not want companies to hide from accountability, which completely abolishing the Impressumspflicht would do.
Also note how I said the GDPR also requires transparency with regard to who processes and stores your data. This doesn't translate to the same requirements that exist for an Impressum but for companies and registered organizations it's enough to make them identifiable and recognizable, especially in combination with the Transparency Register, which is also part of EU law.
You're right legally, but obviously the GDPR is not fully followed -- Big Corpos just ignore it and pay the fine, and small companies can skirt it.
I don't understand your overall point about "data" though. Do you mean for free usage when people accept cookies from their logging, or just for customers of the API since you make an account?
In any case it looks like the FAQ now links to the parent company, but I could have imagined it just being a guy who didn't want to get doxxed or wanted to stay private.
I think being able to make a website or tool or thing and say "hey check this out" and stay anonymous is a key part of the internet, and frankly I don't mind if they make a small amount of money on the side. I know this is probably Ketzerei in Germany but in Anglo countries it's sometimes notoriously hard to track down corporate structure to people and such.
Germany is definitely incredibly pro copyright though so that probably plays a role.
Data about you is your data. The GDPR defines it as such. As long as data can be traced back to you, even through pseudonymization, it remains your data. This includes anything from IP logs to what you did in the app. If it's tracked, that generates data, the data is tied to you, so it's your data. Given that the app invites you to upload pictures, which themselves could be other people's data, it's very relevant to know who is storing, transferring and processing it and for what purposes.
> Germany is definitely incredibly pro copyright though so that probably plays a role.
Sure, to some degree. I'd also like to believe that we have a heightened cultural awareness of the dangers of governments and corporations having access to personal information when things go south. The biggest civil control mechanism of the East German government was what at the time would have been considered an excessive amount of data collection about anyone even remotely suspicious of being critical of the state (and anyone affiliated with them). And prior to that the NSDAP used intricate record keeping to identify "Jews" and suspected enemies of the state. It doesn't matter if it's a corporation that has the data or the government because fascism doesn't make this distinction. So the only way to protect data is to have full transparency over who has it and why.
Edit: If you get contacted by a foreign intelligence agency attempting to blackmail you, please report them to https://www.inscom.army.mil/isalute/ and also to 1-800-CALL-FBI (1-800-225-5324).
Yes but why North Korea? This sounds like US 3 letter propaganda from a decade ago
It would be China looking to Exfiltrate IP/Data as that data war has already begun. See : Azure keys.
If the images posted are the best they can do, then i have some bad news from them
assuming you have the dexterity to hold each of your 10 fingers in one of 3 different positions—e.g.: up, down and crooked—you could technically count up to 3^10 or 59,049
if you mastered that, you could add a fourth or even fifth position. 4^10 is just over 10^6, so you could hold any 6 digits, and 5^10 is just under 10^7, so almost any 7 digits
you can compress really heavily using high bases, but for it to make practical sense the atomic medium of storage has to have that number of states
if you just convert binary to a high base and back again, unsurprisingly the space used works out exactly the same, as the number of symbols used to represent a number may be fewer, but the length of the symbols themselves balances it out exactly
e.g. in base 1000 you may be able to represent the number 999 with a single symbol, but you need 1000 unique symbols to show every number up to 999, each of which of course takes up significantly more digits than a bit; however, if you had a storage cell that could reliably store and display 1000 unique states, then it would be a different matter. whether that's possible or would/could even take up less physical space than ~log_2(1000) binary cells I do not know
A mermaid with plastic looking skin, and badly rendered ocean water in background.
https://generated.photos/human-generator/64d6dde03af7f90007c...
It’s laughable to call these “hyperrealistic”.
If there is no copyright in AI-generated images, then how can they possibly enforce this?
Additionally, even if there's no copyright, the terms of service may still apply separately (see OpenAI disallowing training a competitor model on output from OpenAI models)
I would say it doesnt look good at the moment for to try and enforce ownership of something ai generated, it would be an uphill battle, and the default/null position would be that the art is free to use, and unprotected by government.
No, they haven't.
They've said that if the only human input is a text prompt, then it lacks the required human creativity to be eligible for copyright protection.
Aren't contract clauses that relate to the distribution of material preempted by the copyright act?
For example you and I could enter into a contract for me to use AI to generate something that is not copyrightable from data you provide and give you a copy of that thing. There would in general be no legal problem if the contract included restrictions on what you could do with that thing, including restrictions on distributing it.
Part of the quid pro quo of a contract can be one party giving up a right to do something that they would normally have a right to do.
Now suppose the contract did allow you to make and distribute copies as part of your product. Someone else starts making copies of those copies you distributed and distributing those copies.
There is no contract between me and that person, so I would not be able to stop them. I've got no contract with them, and the thing is not copyrighted, so there's nothing that prevents them from copying it.
Generally, no. It's possible for there to be interactions in some cases, but the Copyright Act wouldn't generally preempt contract terms. (Its closer to the other way around, in that—to the extent copyright rights exist that could otherwise be enforced—a relevant contract will generally limit enforcement and recover to breach of contract rather than bare copyright action.)
The only real problem here is that the original creators of the art that these AIs were trained on didn’t consent to this type of use and aren’t getting any kind of attribution or payment. If they were recognized and compensated, there’d be really nothing to talk about here - any work could be copyrighted, with whatever derivative status the AI bakes in.
however if there were large nations in which the law was still in flux or unclear, or one wanted to generate new images on the fly without fear of rate-limiting or refusal of service, then one would potentially wish to work out an arrangement.
We don't have precedent here. Whether a person using a website with a generative AI tool counts as having a non-human creator isn't clear, and it seems to me like the answer is that it does have a human creator. Using a horse-hair brush to paint a painting doesn't mean that the painting was created by a horse and isn't subject to copyright. We'll have to find out eventually whether over a dozen settings, some with a gazillion options, and multiple freeform inputs counts as 'not created by a human'.
The horse hair example is nonsense. One might argue that artists take inspiration from other artists to make the argument that what the AI is doing is fine. But the ai is actually only capable of blending what it’s been trained on, whereas an artist is not similarly limited. And this is how the horse hair sample is stupid.
No, there was a recent case where someone tried to claim an AI as an author for copyright and themselves as owner via work for hire, where it was ruled invalid because AI can’t be an author under copyright law; the ruling was explicit that it was not addressing copyrightability by humans of images they create using an AI generator as a tool, only the claim of copyright with AI as the author.
No.
> One might argue that artists take inspiration from other artists to make the argument that what the AI is doing is fine.
This doesn't seem to address what I took to be the relevant part of IP law - that non-human authors don't create copyrighted works. It was a reductio ad absurdum for minimal non-human involvement. It's probably not the case that a monkey stealing your camera and taking a selfie creates a copyrighted work. It's probably the case that a frog triggering a motion sensor you set up for nature photography does. It's certain painting normally with a horsehair brush does.
Your remarks seem to make some sort of moral appeal, but I'm not sure how it ties into the legal concerns I thought was being raised.
> the ai is actually only capable of blending what it’s been trained on, whereas an artist is not similarly limited.
I'm not sure what "blending" means here or what the actual theories of generative art ML systems and of humans here are. To call what the former do "blending" requires such a broad definition I can't tell you if humans are blending as well (at least some of the time, at least materially) when creating works.
(If you're thinking of the recent court case, no, that was unrelated; some guy was trying to pull a stunt and the court did not actually rule on the thing you think they did.)
How could they enforce it? For example, by embedding a steganographic tag, by which human rights activists will be able to identify the author of the picture.
[1] https://commons.wikimedia.org/wiki/Commons:Photographs_of_id...
Edit to add: It's not fast, it's showing you repeats of stuff of already made in the first try. Which is probably why I got 5 men in tight pink shirts eating cake in a row. ???
Trying to 'wing it' with engineers doing what designers should be doing is a bad enough when you're just making regular interfaces, but when you're trying to sell a commercial art product, you need people with subject matter expertise. No matter how cool the technology is, and no matter how well it theoretically serves a commercial art customer base, if you're selling art, it's going to be critiqued as such. Hope you've got a thick skin.
Radiation levels??
Also, what the hell was this trained on?!?
I like the UI functionality though. easy to dial in what you’re looking for
That is actually a hilarious issue.
Its not, I got rolled-but-long white denim pants, white shirt, white tie, and white jacket, with white deck-ish shoes but selecting “Formal” on the clothing tab and entering “Clothing: white-tie formalwear”.
But, yeah, there is a definite denim bias.
I don’t understand why none of these generators want to make a man with body hair. I specified Armenian and Armenians are notoriously hairy.
Edit: this is if I specify “very hairy”. Note that for a man, there is a bit of hair in the usual places but still far short of “very hairy” IMO. https://generated.photos/human-generator/64e67aa38448b800095... (And the hair rendering is bad, AI still doesn’t understand directionality; there’s a very common pattern it should know, especially on the chest)
The need to track people is strong.
This has widely been seen as something of a problem, environmentally.
In short, as a tool for an art director or illustrator trying to avoid the expense of a model or a photo shoot, it's fuckin' useless.
It's harder to do at inference time without training, but I wouldn't assume it'll be impossible forever, especially with the existence of ControlNet.
Is that how the latent space works though; Like if it's a 300-dim vector, is the face at locations 0-10?
But Poser is 28 years old now, I guess it doesn't attract sweet VC money.
In my mind, it would be smarter to use the AI to tune the 3D model, and then let a rendering algorithm like this generate the final image.
Edit: lol https://generated.photos/human-generator/64db2561ba3ed6000ca...
I doubt everyone is going to want to have six mangled fingers on one hand and a claw on the other.
Tomorrow's "Now", maybe.
We detected that generated image contains nude content. Try changing parameters.
Not sure what to make of this, but it feels wrong, somehow?
Edit: This was the prompt it generated for me on page load: "Minerva McGonagall in Hogwarts, wearing Hogwarts robe and witch hat" - https://generated.photos/human-generator/64e650a39563e6000e0....
I would be more surprised to get SFW results from that prompt, considering the result would be based on more heavily regulated (less common) photographs of minors.
The site lists all of these things you can do, but are those things people needed or wanted? Is the idea to replace stock photography of people?
https://generated.photos/human-generator/64d552c85263da00077...
The tech is super interesting but this space is also a good example of getting the last 1% working is going to be a lot harder than getting the first 99% to work.
For some reason, all the images I have generated so far are unexpectedly sexualized even I when i try to give it completely benign prompts. I don't think it's just me. :-/
Cubic feet per minute?
Seems like you'd want huge amounts of photos of people (abundant publicly, though not necessarily free for use) with huge amounts of tagging for assorted clothes, body styles, accessories, nationalities, scene types (arduous?).
Ethnicity: American
What does that mean in latent space and does this mostly represent training bias?
> We detected that generated image contains nude content. Try changing parameters.
I clicked the female generation, and got a porn model posing nude. Without any provided guidance other than the clickable buttons.
Or maybe a people trying clothes on virtually.
Creating/publicising a tool that winks at these issues (consider the difference between the poses offered for 'male' and 'female' bodies) but does nothing to mitigate them - and a lot to enable them - is irresponsible at best.
Only when published, and when put into context of the entire work, could a creation deemed harmful. A tool should not, for example, prevent you from making a bunch of images with ominous poses, from which you select one to use with an article that discusses the history of ominous poses.
Tool makes do err on the side of caution all the time - we **** out passwords so users don't share them as easily, we put safety catches in secateurs. "Build in safeguards against the obvious issues" is a basic design step.
- assuming you found an egregious issue, do you also double down on maintaining that the tool is 'deliberately designed to make negative uses easy'? How so?
- I disagree with the 'safety catches' metaphor and would offer the 'hammer' metaphor instead.
- Actually, with the rapid development in this field I expect that anyone will be able to locally prompt for any content, even movies, soon, limited only by people's taste and imagination; with this realization I don't think I will follow up on this discussion that will surely be outdated in a minute.
I don't disagree with you entirely, but I still have the feeling like this will make a pretty good epitaph for humanity some day.
Everything from realistic faces to realistic posts. We tried to make it as ethical as possible in multiple ways. But ultimately it is designed to spur conversation on topics that need more kickstarting engagement…
All these AI content generators are still early stage. So it's kind of wild west for the time being.
First cars were what? Horse carriage with an engine duct taped onto it? Only when they became more numerous, things like traffic rules, reliable brakes & steering etc became important.
We're in the engine-with-wheels stage. Have fun, be happy.
One cookie looks like it just records whether or not a tooltip that they want to show to first time users has been shown. The other two appear to be some kind of session cookies.
They might count as strictly necessary cookies.
All the best,
-HG
Actually NSFW, not safe for sanity. That's...not how body parts work:
https://generated.photos/human-generator/64e644f39c8c0400108...
Prompt was "young woman with tattoos in miniskirt" really nothing crazy there. But perhaps the latent space with that particular pose is particularly raunchy.
Not quite right. I am, however, impressed that the fingers are generally "mostly" correct.
She has a second set of boobs where her hips should be! That's not evolutionarily advantageous!
Although not my kind of thing, I'm pretty sure some people would select for that (in real life) if given the option. ;)
I wonder if their pose detection/interpolation struggles for rarer poses, eg "kneeling with legs splayed leaning forward" is quite specific in saucy contexts and fairly sparse in more typical model shoots, so the manifold gets a bit holey, and overlaps with similar poses like one knee up, one hand forward.
Also, it's way too easy to make something that looks like (or basically is) child porn with this tool. I chose Adult but something else keeps triggering it to generate a child-like face like in the image I posted. And as you showed, it's easy to get accidental NSFW pics.
I just let it generate a random woman with no prompt, and it gave me a pretty good result, except there is a mask on the face and literally bloody nude boobs; https://generated.photos/human-generator/64d67874568faa0007a...
edit; I just realized it put in a default prompt
Some are really bad at filtering. Kawaix is particularly terrible at it, because they are new, while mage have upped their game a lot but had many months to do so.
It feels like 2000' again, and it's the wild west.
Plus when you have a horde of teenagers having a whole summer to try prompts from their bed, you get serious pen testing sessions.
https://github.com/stoically/temporary-containers/issues/618
I use home and work profiles on my laptop for instance, works really well.
If this service is hell bent on raping your privacy, they will have to limit their offerings to mostly those living in dictatorships and immature democracies.