Halloween costumes by the neural net GPT-2
4 min read

Halloween costumes by the neural net GPT-2

Halloween costumes by the neural net GPT-2
sentient stone, skunk in a moose costume, eight ball of wrath

In my opinion, one of the best applications of neural networks is for generating Halloween costumes. Thanks to a dataset of over 7,100 costumes crowdsourced from readers of this blog, I’ve been able to generate Halloween costumes with progressively more powerful neural networks. In 2017, I used char-rnn, which learned to generate costumes starting from no knowledge of English (Statue of Pizza, the Fail Witch, Spartan Gandalf, and Professor Panda were some of its inventions). In 2018, I used textgen-rnn, also training from scratch, and teamed up with the New York Times to illustrate the costumes (some of my favorites were Sexy Wizard and Ruth Bader Hat Guy).

Now, as of 2019, there are much more powerful text-generating neural nets around. One of these is GPT-2, trained by OpenAI on a huge dataset of text from the internet. Using the connections it’s gleaned from this huge general dataset, GPT-2 can generate recognizable (if often weird) lists, mushrooms, British snacks, crochet patterns, and even a to-do list for a horrible goose.

So, I trained the 355-M size of GPT-2 (the largest I can currently finetune for free via Max Woolf’s collab notebook)

GPT-2 is good at costumes. Many of its inventions could easily have come from the training data. In fact, the neural net did tend to memorize the training data and repeat it back to me - technically this is what I asked for when I asked it to predict the training data. (The neural net is trying to give me exactly what I ask for, which isn’t necessarily exactly what I want.) I was using a handy script to filter out duplicates (thanks to John Tebbutt), and even so I had to check several of these to make sure they weren’t near copies of the training data. My previous Halloween costume generators would not have been smart enough to come up with things like “jackalope” or “Carl Sagan”, but GPT-2 has seen these words used online in similar contexts to things that ARE in the training data, and it makes the connection.

vampire rock, gothy giraffe, battle worm
Gothy Terminator
jackalope
vampire cat
Eye of Sauron
incognito llama
space cow
Vampire Rock
Scooby Gadget
a raised eyebrow
Battle worm
Mastodon
Swamp girl
Carl Sagan
A space squirrel
walking carpet
Frizzle the witch
Cleopatra on vacation
gothy giraffe
Sexy Lego Batman skeleton

Oh yes, the sexy characters. The neural net definitely picked that up from the training data, and innovated admirably, bringing in words that it knew from the internet (barnacle, groundhog, and bunsen burner were not in the list of Halloween costumes), and adding a sexy twist. This is impressive (if somewhat horrifying) work. None of these were in its training data, but I wouldn’t be surprised if some of them exist.

sexy hummingbird, sexy flying dutchman, sexy barnacle
Burlesque Horse
Sexy Bulldog
Sexy Egg
Sexy Parsley
Sexy Barnacle
Sexy Walrus
sexy locust
Sexy Titmouse
Sexy Hummingbird
Sweet Potato Burlesque
Sexy Groundhog
Sexy bitcoin
Sexy DNA
Sexy Rubber Duck
Sexy Bunsen burner
Butt-Monks
Sexy gingerbread man
Sexy Flying Dutchman
Sexy Chimneywatcher
Sexy Flames Of War
Sexy English Tea Party

And the neural net was pretty good at designing identifiable characters, even if they are a bit on the weird side.

gingerbread man guinea pig, ghost in a packet of potato chips, cozy coconut
A spangled Auroch manatee
M. Bison the Clown Prince of Darkness
Gingerbread Man guinea pig
Skin Fairy
sentient stone
fast food bald eagle
Fairy root vegetable
Ghost in a packet of potato chips
cozy coconut
Kelpie the mage
Crochet monster
Walrus rider
Star skunk
Slytherin AI priest
A skunk in a moose suit
Semi-molten Kool Aid Man
Time Lord Power Ranger
The Power Dinosaur
Space Oystermonger
Deadly Snow Monkey
An evil cupcake
basic plumber’s equine
Spooky mother hen
The Bozo the Destroyer
Eight Ball of Wrath
Ursula, Queen of the Fart Science
A poker player in possession of an onion

There are hints, though, that this is the work of an AI rather than the work of someone who understands what costumes are and how they work. These, for example, take somewhat ordinary costume concepts and then make them unnecessarily difficult.

Batman on egg
Vampire in hot tub
A Hidden Jesus Statue
Zombie ice cream cone
penguin as a Newt
A wizard encased in a icicle
Zombie fisherman on a quest
Computer generated horse(?)
telephone that accepts up to 4 numbers
Third Eye Blind Photographed By Dorothy
Zombie fisherman w/ lady diegrove tied around foot

And the following costumes are clearly the product of a glitchy AI:

pajamas made of wood and spiders, world's nicest fart, list of leg parts
Meat Belt
Eyeballed Balloon Men
Green beans in bun
10,000 Hands
Favorite Caterpillar
The Oatmeal Tree
102 SNOWBALLS in a basket
Pie and Jell-O
List of leg parts
world´s nicest fart
Pineapple wrapped sasquatch
Is it a Snake, a Watermelon, or a Bush?
Putting Turtles on Decor
Fish tank ‘n chair
ROBO-ACCIDENT
pajamas made of wood and spiders
Ssssssssssexy SSSssssssstinky Ssssssssssssexy ssssssssssssssssexy
setup 9 × 11 party trick
Smagma Monster
Commentary couldn’t be heard over the squawking of clocks
Poltergeist might be entertaining, but he’s harder to read in Hungarian
Cereal Implanting Device
blueberry sipping fizzy pop with eyes of ice
blueberry sipping fizzy pop with fake blood on it
A sarcastic, racist noble using progressively tinier body parts as a human shield

The above costumes are all from temperature 1.2; I also tried a higher temperature setting, but the generated costumes were at an expert level of chaos (I would like to see someone attempt to go as “hypnopotamus embroidered death”). AI Weirdness supporters get them as bonus content! Or become a free subscriber to get new AI Weirdness posts in your inbox.

Subscribe now

You can order my book You Look Like a Thing and I Love You! It’s out November 5 2019.

Amazon - Barnes & Noble - Indiebound - Tattered Cover - Powell’s