These Birds Do Not Exist



Great Northern Oyster Cracker



The Tawny-Breasted Fishmonger



Lester's Finch



The Forbidden Duck


The Lesser Bobber



The Rabid Wood-chipper



Henson's Fractal



The Surly Drunkard


Spectacled Ringwraith



Headless Horseman



Eastern Grump



That-Which-Sings


Turner's Hooch



Red-Breasted Eye-Poker



Giant-Headed Bug-Stabber



The False Moffat


Humpbacked Quail



Captain Falcon



The Interrupter



Warped Dangler


Biblically Accurate Grouse 



Monochrome Reedwacker



The Worst Shrek



Blue-Caped Buddy


The Flying Nudibranch (shown with chick)



Jerry's Neighbor



Lovely Lady



The Hostile Witness


The Sharpest Duck



Common Mussel Puncher



Marbled Slab



Hooting Balloon


Common Benchpress



Flagrant Scofflaw



Rusted Rooter



Wooden Spoon and False Wooden Spoon


The Freelance Bartender



The Whole Wheat Breadbox



The Burrowing Evildoer



The Flammable Finch



The Nasty Janet



The Painted Pest



The Dominant Dingus



The Bastard


The Belter



Cloud Strangler



Sailor's Moon



The Flightless Biff


Blue-Breasted Creek



Sudden Turkey



Behold Thine Own Sins Made Manifest Before You



Red Bunt


Tony's Hawk



Donedone, the past tense of Dodo



Dwayne The Flock Johnson







Definition of Terms

I'm saving a more detailed policy statement about AI image generators for a future post. For now, I just wanted to preserve my notes from earlier this year in case Twitter goes down. Note that I'm careful to use the term "image generator" rather than "art generator." I call the stuff that AI poops out "images" and leave it at that. I graduated from art school. I've had enough "what is art?" debate to last years. 

The Beginning

In the late winter of 2021, I first started playing around with AI image generators of all kinds. I was inspired by the light-hearted fun of Janelle Shane's AI Weirdness and wanted to explore this new tool myself. Hard to believe it now, but at the time I considered these image generators more like a weird toy than any existential threat to art itself. 

My immediate concern was the training corpus being so hidden to the front-end user. You had no idea how the AI was trained or what it was trained on. Still, with only a little work, you could find the famous photos that lurked somewhere in the training corpus.

https://www.artbreeder.com/beta/image/ed9a983624520d18e6ed
https://www.artbreeder.com/beta/image/ed9a983624520d18e6ed


I started looking at Colaboratory notebooks that I could train myself. I figured that if I knew where the images came from, I could be a little less nervous about accidentally infringing on a trademark. I trained an early version of LookingGlass on this curated collection of public domain ornithology illustrations sourced from the Biodiversity Heritage Library. 



The results are janky as hell, but also kind of hilarious. Since birds often appear in the context of flowers, leave, and branches, there were often some odd blends of all three. I'm surprised it didn't occur more often, to be honest. It's pretty good at knowing leaves and wings go together. The beaks and heads give it trouble though.



Given the widely varying morphology, I feel ill-equipped for judging the success of the AI's training. I just have a gut sense of what's "plausibly bird." The best I can do is like "Uh... Beak? Yeah, somewhere. Wings? At least 2. Funky plumage? Yep. Plausibly bird. Next!"


The Origins of STORK

I can have a juvenile sense of humor, so I started having fun making up names for these birds. Eventually I compiled some of these images into a jigsaw puzzle and other merch on my Society6 shop. 
 

I created a fictional world behind the images: The Society of Theoretical Ornithology, Research, and Knowledge (STORK) documented the rarest species of avians around the world, noting their behavior and ecological environs. Anyone who has followed me since publishing Belle of the Ball knows how much I enjoy writing lists of ridiculous names and short absurd blurbs. This was the perfect project to indulge that habit.



On a certain archipelago, the fertile phase of the "Nest fungus" mimics bird-shapes. Local birds investigate these mushrooms, dropping fresh guano for on the nutrient-poor rocks. Burred spores then attach to the bird's feathers to travel from island to island.


A selection of symbiotic fungii and their partner bird species. Note the dappling along the stalks and caps matches the partner's plumage. The birds seem to benefit from this relationship by coating their nests with potent scent that repels predators.


In due course, the line between fungus and avian begins to blur, with young chicks being indiscernible from their fungal counterparts. 



Branching Experiments

I did experiment with new settings that smoothed out the pixels, but the results felt more like vintage 1980s hotel wall art. It's charming in its own way, I suppose, but it didn't hold the same appeal to me as the more scientific-skewing illustrations.






I also tried mammals, fish, and chimera, but the results also didn't really have much appeal to me. Mammals in particular didn't have the same variety of color, but at the same time were too varied in morphology. So I got a lot of tripedal furry creatures with bland, greyish fur.





Perhaps 1 in every 50 outputs were completely alien. I think the preponderance of legs, trunks, and tails in the training data confused the AI, so it just made these tentacled horrors.




Occasionally it was just a long snoot.



And that's basically it. I had no greater aim for this experiment than to have some fun and make some weird mutants. 

Faith Plummets

I tried using newer generators as they arose in mid-2022, but none had the homespun charm of the AI that I had trained myself. The more advanced they generators became, the less weird the birds. Furthermore, the newest generators presented even more conundrums and egotism that soured me on the whole notion of even toying with AI anymore. These results were from the Dall-E generator in early July.






Even the AI I trained myself wasn't without its own problems. I was naive and didn't understand at the time that the LookingGlassAI already had some pre-training before I added my own dataset. There were no humans in my data set, yet LookingGlassAI occasionally insisted on trying to make human forms. Some of these are actually quite gothic and haunting. Nevertheless, it made clear that even this seemingly more-ethical AI option wasn't safe from risk of infringement.





This became most evident when stock photo watermarks started appearing in some of my results! 




Conclusion?

And so my last project using AI was "These Books Do Not Exist" jigsaw puzzle. I'm still mulling over what I will do with AI next, if anything. 



I've saved a bunch of LookingGlassAI outputs in this folder for posterity. Who knows what I'll do with them next, if anything. As I've come to realize, no AI-generated image is without legal risk, even if I make a transformative work like my posters and jigsaw puzzles. 

So for now I'm happy with the weird little birds that I made over a few weeks in January 2022. Making up names was fun. I hope this silly venture brings a little joy to your day as well.

Comments

Post a Comment

Popular posts from this blog

5 Graphic Design and Typography Tips for your Card Game

Image
Chris Farrell of Illuminating Games just wrote a thorough critique of card games setting their body text to be so small that they cannot be read at arm's length, let alone across the table. Here's an excerpt, but the whole thing is worth a read for any would-be card game designers out there. (Myself included.) Break out your copy of the base set of Dominion, and look at the Chapel. This has a text box roughly 3.5cm by 4.5cm. It's got a single line of text. That line of text is 1 (one) millimeter high. 1mm! For me, it's only clearly readable at half an arm's length even in the bright light of day. (...) Compare this to a more sane game like Glory to Rome, where the font size is 2-3 times as large (text is 2mm high, keywords 3mm and usually highlighted). I can generally read Glory to Rome cards across the table, and can certainly see the important keywords. To heap insult upon injury, not only does Glory to Rome have far more legible text than Dominion it also has
Read more

One Thing to Avoid in Game Design

Image
When game designer Paul Peterson was interviewed for the upcoming Titans Series of games from Calliope Games, there was one segment at 1:55 that really caught my attention (emphasis mine): "A couple years ago I was teaching people how to play Guillotine and I hadn't played in a while. The first thing I did was say, 'The first thing you gonna do when you open your game is you look through this deck and you take this card.' I pulled out the Callous Guards card and threw it over my shoulder. Everybody laughed. I said, ' That's the biggest mistake I ever made as a game designer in my life.'" Oof. Those are strong words. What's the big deal? For context, the game Guillotine is all about collecting the heads of French nobles during the revolution. There is only one guillotine, with a line of nobles waiting for their grim fate. On your turn, you play action cards to change the order of the line, then take the Noble card from the front of the
Read more

Belle of the Ball Guest Name Generator

Image
Here's a fun little toy to tide you over while Belle of the Ball's is still in playtesting. Come up with your own name for a Belle of the Ball guest! Start with a first name from the first table. Then add a surname by combining a prefix and suffix from the second table. Bang! Now you're one classy Dental Untercaw or Chive Pimplesack!
Read more