What I learned about inclusivity, male gaze and NFT promotion doing a project about boobs.
There was only one thing to do: “Teach an AI to draw boobs in all shapes, sizes and colours #freethenipple”
TL;DR
- Taught the computer to draw female chests.
- Created an NFT project somewhere between collectible avatars and nude art with a deeper message.
- Tried to make people think about double standards, prudishness & implications of AI nudes.
- Learned: Internet boob data heavily biased towards larger white breasts.
- Confirmed: Nothing wrong with boob appreciation. However, as an artist beware that it could undermine the message you try to convey.
- Learned the hard way what works and what doesn’t work when promoting your NFTs.
- Donated part of the profits to pink ribbon
Why I am writing this:
I want to share what I learned while doing this project. And I also feel the urge to explain how much work and consideration went into this project. Since it turned out looking quite good, some people assumed I just applied a fancy filter to existing boob images. Even though that is quite the compliment, I still like to share that there is a lot more behind this project.
How it started
If you are still reading, awesome. Keep going! This project started after some conversations I had about censorship of female nipples (but never male nipples), the taboo on public breastfeeding and other double standards. An interesting thought came to mind: isn’t this kind of behaviour contributing to the sexualisation and objectification of boobs? I mean, in the end even though they look beautiful, they are just boobs and 50% of the world population has them. They occur in all shapes and sizes.
Being a mediocre programmer with a vivid imagination and an interest in NFTs there was really only one thing to do: teach an AI to draw boobs in all shapes, sizes and colours #freethenipple
I named the project: The Treasure Chests. I was going for something somewhere between avatar collectibles and nude art with a deeper message. My goal was to make people think about two things:
- First, prudishness in society and double standards when it comes to (nipple) censorship. For example; breast feeding vs. sexy pictures on social media (#freethenipple)
- Second, what are the implications of a computer being able to generate nudes? For the adult industry? For beauty standards in general?
- And to be perfectly honest, apart from the above and being interested in anything related to blockchain, the insane amounts of money spent in the world of NFTs are also quite intriguing.
The steps to AI generated breasts
Step 1: I had to collect a training set first. Images of breasts the computer can learn from. There’s more than enough of that. However, you don’t want to get them from just any website. Preferably, you want to use websites/forums that are monitored to reduce the risk of illegal material slipping through. Using and tweaking some existing code I could scrape specific threads/channels which I trusted. Another solution I applied was adding a lot of 3D animated images.
Step 2: Since I don’t work at google and don’t have access to insane computer power (GPU’s), I couldn’t just use all these images (100k +). I had to create a smaller high quality dataset. I had to get creative. My solution was to tweak the code of an existing nudity filter which helped me select the best images and crop them in such a way that they only contained the chest area (so no faces or limbs). Using the nudity filter I could also exclude any (parts of) images that where way too graphic for this project.
Step 3: Then a problem occurred. Not surprisingly to be honest, the data was heavily biased towards larger white boobs. There was some diversity in the training set but way too little to have an impact on the final output. Ouch.. this was exactly the opposite of what I was going for. I wanted breasts in all shapes, sizes and colours.
The problem of biased data is (imo) the biggest challenges in applied AI. I can recommend the documentary coded bias by Joy Buolamwini if you want to learn more about this.
For me this meant, I had to get creative again (BTW I won’t bother you with all the technical challenges I faced during this project, but if your curious, just contact me).
Step 4: I tried various sampling techniques but with few results. There was just too much bias in the data. I also wondered if I should just leave the bias present to show how biased the data was. This problem exists much more problematically with face recognition algorithms. With huge implications! (contact me for interesting reads on this topic). It could therefore be interesting to address this problem in a seemingly innocent boob project by leaving the bias in. In the end I decided to go another way….
My plan was, after the generation of realistic boobs to use another AI technique called style transfer. Such an algorithm is taught to transfer the style of one image to another. I had many different existing paintings in mind, being human and biased, I easily turned to painters like Picasso, Vermeer and van Gogh. However, at one point I decided to include traditional and modern African paintings/painters. And suddenly, my outputs became much more diverse and beautiful paintings.
Critical note: Is this method of manually introducing diversity ideal? Far from it. Ideally I would have been able to teach an AI diversity. I decided having ‘a diverse final result’ outweighs this issue. However, I am interested to hear other opinions on the matter.
I also really liked the fact that my AI generated chests didn’t show perfect symmetry, as can be seen below. However, the bias towards larger breasts still exists.
Step 5: The final step was minting. I was quite happy with the results and entered opensea and twitter with confidence. Well, that was a bit of a disappointment. I thought the NFT world would be a lot easier. But just like the real world, getting your work noticed is bloody difficult. I made all the rookie mistakes (and am still making them): Gifting my work for free to influencers (oops), DM’ing, spamming. Sometimes I felt a bit frustrated. Didn’t those people see the work I put into this? Is this just a big failure? Is my work even worse than some of the low effort avatar successes? Wat kept me going was the handful of people who did get the message. There were also some purchases in the category; “Boobs, I like boobs, I buy boobs”. Nothing wrong with that. In the end, the project is also an homage to boobs. If you like them, buy them 😊.
Conclusion
So this is where I am currently at, somewhere in step 5. Sales are slowly going up. But the positive feedback I get every now and then, from people who got the message is the best. I really felt I had to write this small article to explain the process and intentions of this project. Hope you found this interesting and are still reading. For questions, feel free to contact me.
Cheers.
Additional remark: I sometimes refer to the outputs as ‘female chests’. But actually, we don’t know their genders and there is no way of telling. Do with this information what you want.
Additional additional remark: The original idea also included using the generated breasts as tiles in a mosaic. Using the generated breasts to give a whole new meaning/interpretation to the complete image. I did mint some examples of this. They can be found here.
Final notes:
- Peortsitid also made a lovely collection originating from a similair idea. This project is called 8-Bit Tits. This collection inspired me to start donating to pink ribbon with my collection as well
- Special thanks to all buyers of pink ribbon items!
