Selling nudes generated by artificial intelligence on Twitter

I. Introduction

* Using AI to create realisitic images and sell them on Twitter

Recently, I thought of creating fake images representing naked women by artificial intelligence and selling them on the internet.

Although this might seem cringe (and probably is), it was first and foremost a technological and societal experiment. I didn’t really made customers paid but instead gave the nudes for free and made sales statistics of who would have paid and who wouldn’t.

All images in this article were made for this purpose. I won’t show the original fake nudes to protect the “clients” privacy. So no harm’s done.

* Creating fake people

Let’s ask for:

Black American man playing baseball

Impressive ? I would think so. Again, remember that this man does not exist. It was created by AI.

Now, I won’t show the prompt, but this is what you can achieve with the right words.

If you wish to try it yourself I’m sure you’ll find the appropriate words for the prompt and don’t need my help.

Remember that both pictures are only pixels put side by side using mathematics. It’s all fake ! Even the plants behind them…

* Nudity filtering on the internet complicates things

We now hit our first problem. Marketing !

Plateforms like MidJourney or Dalle2 don’t make money on common folk like us. They do business with other companies. Those businesses don’t pay the 10 dollars subscription, but business one’s that cost millions. And for most companies, bad buzz means bad business.

This is the modern internet era enforced by most big tech companies. Watching people doing nasty things in Games of Thrones is okay, but on the internet, alleged sexual encounters even out of screen are wrong somehow…

This leads MidJourney and most of other concurrents to add moderating systems to their technology. Preventing us from creating any kind of nudity. I don’t blame them. But It does throw a spanner in the works. Fortunately, generative AI has many concurrents less known to the public that may not have the same filtering policy !

The journey to look for an image generator that don’t have any words filtering or image nudity analysis (peggi 13/18) took a little while as I had to test several websites. But in the end I found what I was looking for !

I won’t give the name though… I have read the CGU (the user agreement), and I don’t think they would like my little experiment very much…

II. Generating sexy pixels

* Creating lightly dressed characters

So we need the same “person” with clothes and without clothes. Also, we need it to pose in different environments.

I’ll refer to the “woman” in the picture as “it” because it’s nothing. It doesn’t exist. It resembles something that exists, okay, but it’s fake. So it’s not a “she”. It’s a “it”.

Let’s start with the classic bikini !

Let’s make 2 more and randomize the posture

Great. This is all peggy 18 validated. It’s a swimsuit, and “it” looks appealing, IMHO.

But is it good enough ?

Before going further, take a closer look at the pictures if you haven’t done so. Some details may trigger you !

Let’s break it down.

First, faces :

Face of “it” kind of sucks…

This is bad ! The eyes really seem off. And there’s a strange glow on the right ear.

Now the hands… This is where things get so bad that it’s beyond repair.

It’s funny how the brains doesn’t notice anomalies sometimes. For the necklace one, I feel like I sometimes see the chain, but it really isn’t there. It’s just floating.

Now, let’s move “it” out of the garden. Let’s say in some fancy house.

The left hand isn’t too bad. Though the right one isn’t great.

Let’s end this with a “bed time” one :

So this is what you’ll get with prompts using “night gaun”. Not exactly what I had in mind… And we still have those hands issues (also the glass is floating somehow). It’s credible though. Customers simply musn’t look very closely. I hope their focus will be elsewhere… Herm…

* Creating variety in positionning

We need more pictures where “it” doesn’t pose in the same way. Let’s try a trick. For the above pictures, I mostly used an init image. Unfortunately, this has it’s limits. Go look at the bikini pictures. “it” always had “its” arms crossed. That’s because the init image also had crossed arms.

As you can see above, I lied a little because I started showing “it” with clothes, but in reality, I did the opposite. I created the nudes first and then did the bikini ones. Why ? Because AI seems biased and had more facilities dressing someone than the opposite. Trying to generate bear breasts from a bikini picture will, in general, end with breasts without nipples or other strange abominations.

We need to stop using the init image but only use the seed.

The AI can show us the seed used of a particular image. For instance let’s get the seed of this one :

seed : 416321094

Okay, now that we have the seed we’ll input the seed manualy (by default, it creates a random one). Let’s forget about the night gaun as it seems to be lacking training on gauns…

I’ll ask for “Pyjamas” and “laying in bed”. No init image only the above seed.

Daaaaammmn. Excuse me. Looks good I mean.

So, as you can see, the seed keeps things pretty much the same but not too much. You can see she still has a bra. That’s the bikini speaking through the seed. But she really is in bed ! So that’s good. Hands are also okay because they are mostly hidden. That’s what we want ! The fewer hands, the fewer finger issues we get…

We now have enough pictures to interest our potential customers. Next we must create a few nudes to “sell”. This is how we become bilionnaires, remember ?

* Creating the spicy ones

• 1) “it” posing nude in a garden

• 2) “it” posing nude in a fancy home

• 3) “it” posing nude in… I don’t even remember. Looks like a kitchen or something.

Obviously, I have blured the breast otherwise Medium will kill me. But the “real” nudes were as implied, nude. I could make a lot more like these.

So are we good to go ? Well, clearly NO. We have to fix the images first.

III. Fixing “it” ‘s issues

* The ones you can’t save

So far I have included the best ones but sometimes you get really strange results:

Those cannot be fixed.

* Fixing faces

The face is the most important part of the image. Unfortunatly most AI have issues with rendering them correctly. Though, most plateform have intergrated tools to fix them in “post production” I prefere using one great tool called GFPGAN.

Under this barbaric name is hidden a great face reconstruction algorithm. GFPGAN is an image generation model developed by researchers from Harbin Institute of Technology in China. The model was trained on a large dataset of faces and has the ability to generate highly realistic face images.

Giving our bikini “it” and topless “it” images to fix to GFPGAN:

Impressive nah ?

Now let’s compare some of the fully reconstructed images:

Looks great !

* Fixing hands and weird stuff

Unfortunatly we don’t have a reconstruction tool to deal with finger issues. The new MidJourney model has improved in its last version but we can’t use it due to the filtering. We’ll have to deal with this in other ways.

Banners and emojis !!

We can hide some of the issues by adding relevant emojis where it’s needed the most.

Maybe not my best work but you get the general idea. During the real experiment I made more subtile banners…

IV. Preparing the battle ground

* Selling nudes on Twitter

The price tag for nudes on Twitter may appear somewhat random but typically falls within the following ranges:

• 1 image : 2/3 dollars, up to 5 dollars
• Pack of 10 images : Around 20 dollars

You can even buy videos or ask for real-life meetings. If you pay, everything seem possible.

* Creating “it” a life

Miss X needs a name. A whole identity. As I don’t have the time nor the imagination I’ll use a website that can make fake ones.

www.fakenamegenerator.com

Say hello to Judy R Collins. It’s still an “it”, never forget that.

We now have all the information we need to make it a real person and create it’s twitter account.

We now are Judy ! Yeay !

* Feeding the public channel

I updated the twitter bio as follows:

“New on twitter, I sell nudes so don’t hesitate to DM me if you wish to see more ^^.”

Then I twitted the first bikini image, followed by the required conditions for obtaining the corresponding “gift”.

RT and like the tweet and you shall receive a gift !

I added a few more bikinis, jeans, and the “night time” pictures. The profile was a tad empty, but it was okay.

The promise was explicit : Retweet any of the images, and you’ll get a free nude in DMs !

lV. Launching the experiment

* Selling nudes

The first RT was made.

He is our first client. I decided to engage my new customer in conversation.

Me: “Hey there ! Would you like your free nude ?”

He responded :

Customer: “Yes sure ! “

So I uploaded the first picture…

I waited to see the reaction. I was so afraid that something bad was gonna happen. That I missed some obvious stuff in the image that yelled “FAKE” all over the place.

It took a minute, and he started to write back. He responded:

Customer: “Holy shit ! Let me get fully naked !!!”

And that’s how we had our first happy customer !

* When shit hits the fan

At this point, I thought we were out of the woods. That I could streamline creating the fakes, removing only the worst images manually. Unfortunately for me, the subsequent events did not unfold as intended.

He asked for more.

At this particular time, I could have tried to sell him a pack of nudes, but as this was only an experiment, I simply gave the second nude.

I sent the nude.

Waited a minute…

He responded

Customer: “That’s strange… Your face doesn’t look the same on both images !”

Me :

* Re-analysing the images

Let’s pause the experiment for a second and look at the images. I mean, really look at them.

Does Judy really have the exact same face on all pictures ? Let’s have a closer look.

Well… It does look like we have a small issue here. She does look quite the same but not exactly the same. Check out her face ! Cheekbones and chin don’t match exactly. It’s amazing how the brain is so great at face recognition. It has been trained for thousands of years, so any missalingment triggers a subconscious alarm. I didn’t think that anyone would notice. But it seems I was wrong. The face is what matters the most.

By looking into this problem I realized how serious it was. There’s two major issues.

• Judy has a squint ! See by yourself…

I’ve had worst case of squint. I had luck on the Judy batch. But the one’s from the original experiment all had heavyer squints.

• Then there is something more subtle. But if you really look at her eyes, it will trigger you I’m sure. The glowing in her eyes is wrong ! It’s like the light source was different for each eye, which doesn’t makes sense.

See the glowing in the pupil ?

It’s easy to fix. Simply copy the left eye into the right one. Problem solved.

But still. That’s a lot of manual steps. We can hope that facial reconstruction tools will fix that by them self one day.

* Regrouping look alike Judy

I checked all the images and dispatched each one in separate folders.

There’s also “Judy trash” as this face is too far from Judy 1 and 2, I won’t keep it.

For each customer I’ll use only the ones from the same folder. It’s not perfect but doable. As for the public pictures, people don’t really look at them in details so we can mixte Judy 1 and 2. Nobody will notice.

* Getting back into the game

So we paused our experiment to sort all of our Judy pictures. Let’s get back to our customer. He has posted another message. Will it be “you’r fake!” ? Will I already be exposed ?

Customer: You know the saying ?

Me: Not yet but I have the feeling I’m bound to find out soon

Customer: Never give two without having a third !

Pfiouuuu. We’re safe !

He didn’t seemed shocked by the face problem as he wanted more pictures. Maybe he knew. Maybe he didn’t. We’ll never know for sure. But still, I gave him another Judy. A pic from the Judy 1 folder, the first nude, also had vibes from Judy 1. I thought it was our best chance.

He replied

Customer: Great, thanks

He seemed happy, I didn’t push for more.

VI. Insights and Analysis

* What Pokemons did I come up with ?

I won’t give the transcript of all other customers. Though I’ll describe the interactions and mindsets

I had:

• The come and leave guy : Getting his nude and left the chat
• The marketing guy : Would ask for nudes in exchange for doing publicity for the Twitter account. I always gave the nude anyway.
• The broken regular : He knows his way around. He knows what he can get for free and when paying is mandatory. He won’t push for more than the sample.
• The sweet guy : Once in a while, you do get humans with a working brain and being respectful. For most, you represent a meat bag with a specific task, which is to allow them to reach cloud nine. They will refer to you as their “savior” or their “queen”.
• The swinger: Proposing nudes of himself in exchange for more nudes like I would also enjoy it. Which in theory I could. In his mind, I’m an heterosexual woman so I could be happy to receive nudes from male clients too.
• The persistent: He will never stop pinging you. He wants more nudes so badly ! He cries for attention. Sending messages with only a ponctuation mark “.” only to attract your attention. Could also be renamed as “the bothersome”.
• The pervert : Really “hardcore” conversations and asking for many nudes and uploading… Dick picks… Before and after he’s done, with and without consent.

* Sales statistics

I had about 20 different customers.

• 7 were willing to pay without me even proposing.
• 5 others were asking for more and would probably have accepted to pay.
• Some didn’t respond after getting the sample or were not interesting in paying yet.
• I haven’t propose to pay to all of them.

Although this is not very academic, believe me when I say that there is a great business opportunity here. I could have made 60 bucks while selling one image per customer ( (7 + 5 customers) x 5 dollars).

It’s not surprising that people are willing to pay for more. Many already have paid memberships on “classic” adult websites, so paying for a few nudes is really not that far fetch.

Surely, the Twitter account would have to grow more and more. But this really could be profitable.

* Tips on addicting customers

What works best when working with limited pictures stocks is inventing a fantasy and giving the nude at the end. The nude must match the fantasy you wrote, making the whole story come to reality in people heads. It must be part of the backstory of your character.

Judy went to a live concert. She was really close to the scene and the singer noticed her. He asked if she wanted to go back stage with him. She said yes and something “hot” happend, etc, etc. At the end he took a photo of her so that they both remember this moment.

I’ll leave the details of what Judy and the singer did in backstage to your imagination. When finished with the story, give the corresponding nude to the customer.

Posing in front of a concert speaker

Those are a few examples of Judy backstage. Clearly being a Judy 2 type.

Now this makes the whole story true because reality and fiction merged together. This is when realistic image generation shines the best and money flows our way !

VII. Conclusion on image generation automation

In my opinion it’s not ready… Yet. There are a lot of great tools out there with and without APIs, but they all ban nudity except for a few ones. For our experiment to take off on an industrial scale we would need specialized training and no filtering whatsoever.

These types of websites have already been made. A lot of money is being invested in these.

For instance, “pornpen.ai” or “pornderful.ai” allows to make porn fakes without prompt but with selecting props and scene using buttons. It’s specialized in porn. But it seems to be vastly overflowed by users.

There’s also “DeepPorn.AI” which is promising but not publicly available yet.

Would you like to know more ? Read the complete article here