avatarJim the AI Whisperer

Summary

The author discusses the challenges of generating accurate images of the public domain version of Mickey Mouse, Steamboat Mickey, using AI art generators.

Abstract

The article "Chasing Steamboat Willie: Why AI Art Generators Don’t Mimic the Right Mickey Mouse" explores the author's quest to generate period-accurate images of the public domain version of Mickey Mouse, Steamboat Mickey, using AI art generators. The author highlights the seven distinctive features of Steamboat Mickey that set him apart from later versions, including his black and white color scheme, clothing, lack of gloves, simple eyes, body proportions, personality, and animation style. The author attempts to generate images using ChatGPT, Dall-E 3, and Midjourney but encounters difficulties due to the AI's tendency to default to the most recognized version of Mickey Mouse. The author suggests that the solution lies in custom training an AI on images of Steamboat Mickey while never calling on Mickey by name in the prompt.

Bullet points

  • The author discusses the challenges of generating accurate images of the public domain version of Mickey Mouse, Steamboat Mickey, using AI art generators.
  • Steamboat Mickey has seven distinctive features that set him apart from later versions, including his black and white color scheme, clothing, lack of gloves, simple eyes, body proportions, personality, and animation style.
  • The author attempts to generate images using ChatGPT, Dall-E 3, and Midjourney but encounters difficulties due to the AI's tendency to default to the most recognized version of Mickey Mouse.
  • The author suggests that the solution lies in custom training an AI on images of Steamboat Mickey while never calling on Mickey by name in the prompt.
  • The author plans to create an AI animated short sequel to Steamboat Willie and seeks support from patrons to fund the project.

Chasing Steamboat Willie: Why AI Art Generators Don’t Mimic the Right Mickey Mouse

My quest to get AI to generate vintage, public domain Mickey Mouse

By now you’ll have seen multitudes of Mickey Mouses making the rounds in AI art circles in this brand new year of 2024. I have to apologise for missing the (steam)boat on this one. But as an aficionado of both AI and Disney, and being a bit of a stickler, I haven’t been satisfied with any of my AI images of Mickey. Why? Because there’s a very particular design that entered public domain on New Years, and it’s not the Mickey Mouse you might remember.

Mickey’s revenge on an unsuspecting public. Animation effects added using PhotoVibrance. Try it here!

How to Identify Public Domain Mickey Mouse in 7 Steps

Public domain Mickey debuted in Disney’s 1928 short “Steamboat Willie,” and has seven distinctive features that set him apart from later versions:

  1. Black and White: Steamboat Mickey is rendered in black and white.
  2. Clothing: He is usually seen wearing shorts with buttons, but without the red shorts and large yellow shoes that are standard in later designs.
  3. Gloveless: Unlike the Mickey Mouse that later audiences are familiar with, Steamboat Mickey does not wear gloves.
  4. Eyes: His eyes are like dark saucers: simple, solid black ovals without the complex pupils and detailed expressions seen in future versions.
  5. Proportions: Steamboat Mickey has a longer and leaner body and limbs, compared to the shorter, more rounded figure he would later become.
  6. Personality: The original Steamboat Mickey has a mischievous cruel streak, which is evident in the slapstick humor of “Steamboat Willie.”
  7. Animation Style: The animation style of Steamboat Mickey is less fluid and more ‘rubbery’ than modern animations, reflecting the late 1920s.
Later versions of Mickey lost this spinal elasticity (as simultaneously he became more boring and sanitized)

What about that more rodenty looking Mickey from Plane Crazy?

Plane Crazy Mickey was made a year earlier than Steamboat, and touts a more rudimentary design. But it was actually released in by Walt Disney Studios in 1929, after failing to pick up a distributor the prior year. Plane Crazy Mickey is identifiable by his white eyes, sharp elbows and no shoes.

The Mickey Mouse you’re allowed to mimic

The Mickey that’s up for grabs without a price tag isn’t the cuddly, rosy-cheeked icon plastered across theme parks and childhood memories. No, the public domain character that’s free to sail into any artist’s repertoire is the 1928 version: ‘Steamboat’ Mickey. Very few of the generated images I’ve seen of AI Mickey fit the free bill (or the complimentary ticket, so to speak).

The Evolution of Mickey Mouse — 88th Anniversary. Image courtesy of ikasumefan06.deviantart.com

I don’t think most AI artists realise that the Mickey we’ve got is the rascally rodent that gets angry, kicks animals and has bendy hose arms and legs — not the domestic, sanitized, corporate mascot from the Parks! He’s more like… Steve Buscemi Mickey. And I love him. So much more intriguing.

Satirical piece by my colleague and work BFF Amanda Weston. Gloves blackened in Photoshop.

That’s why I procrastinated for the first few days of the year to write this post. I wanted to generate a period-accurate Mickey Mouse; partly because I’m a perfectionist (particularly where history is involved) and partly out of caution. I get a lot of attention for AI images (they’ve been on ScreenRant, Sky News, GB News and others), so I didn’t want to become a scapegoat. Disney lawyers will come down on someone — and hopefully it’s not me!

Facing the prospect, I had whatever the AI equivalent is of writers block. Generative anxiety? AI overwhelm? However, I still gave it a good old try.

Prompting Mickey Mouse with ChatGPT and Dall-E 3

Another of my online colleagues, the AI expert Christopher Brock, had success using ChatGPT, but you’ll know how mercurial and capricious ChatGPT can be when it comes to edge cases. The usual response is:

“I’m unable to generate images based on the description provided due to content policy restrictions.”

However, reminding ChatGPT of the date can workaround this restriction:

Christopher Brock’s prompting technique to generate a Mickey Mouse image

Another tip that can help get around restrictions is to call the image by class rather than instance. A class in AI image generation refers to a category or type of object, while an instance would be a specific member of that class.

For example, instead of “Mickey”, try referring to the character as “vintage 1928 animation-style anthropomorphic mouse” (with thanks to Amanda Weston). This can often circumnavigate the nuances of content policies.

But I wasn’t satisfied with Dall-E 3's output, which was still anachronistic:

Note Mickey’s gloves, the top hat, and often the presence of the whites of the eyes. Whatever the essence of “Steamboat Mickey” is, this ain't it.

I wanted more specific control, which Dall-E doesn’t offer. Dall-E 3 is a workaday image generator; it’s the most user friendly, but perhaps offers less “earned” results, and the end result is less artistic in terms of control. Despite my recent misgivings at V6 Alpha, Midjourney is my AI of choice.

Prompting Mickey Mouse with Midjourney

Although more aesthetically pleasing, I encountered the same problem with Midjourney. Even when adding the parameter --no gloves, I got unacceptable identifiers of Mickeys that weren’t from Steamboat’s era.

Try as I might, I couldn’t get his saucer-shaped or cartoonish pie eyes:

This actually tells us a couple of interesting things: the AI clusters used by Midjourney keep no real timeline or provenance for images scraped. This would be a very useful feature, especially images entering public domain advance: the ability to selectively use training data only from certain eras.

And secondly, the “thingness” of Mickey is inseparable from his total saturation of pop culture. Mickey is and always has been. He’s utterly ubiquitous, eternal and universal: he’s achieved cultural penetration.

This saturation creates a challenge in generating genuine public domain versions of the character like Steamboat Mickey, whose specific features are less represented in the datasets the AI draws from. The ‘Mickey’ that AI remembers and recreates is the one that’s been reinforced through decades of films, merchandise, and theme parks — not the troublemaker of the 1920s.

The Mouse that became an elephant

The fascinating, unintentional discovery is that Steamboat Mickey is hard to prompt accurately. AI defaults to the most recognized Mickey — the friendly face of the Disney brand. It’s hard to prompt him accurately because of the wealth of later images out there in popular culture.

I believe we’ll see this more and more in the AI art world, especially as corporations like Coca Cola start to enter the fray with branded assets:

Mickey has become an archetype, so deeply ingrained in the lexicon that when AI reaches into the collective consciousness, it pulls out the version that has been imprinted in our minds through countless iterations and decades of branding. Mickey the behemoth overshadows his origins.

Steamboat Mickey in the grand and mythic form of Moby Dick.

Mickey Mouse is going to be like Moby Dick — a white whale of the public domain: much sought after, rarely seen, and an obsession for those of us who chart the oceans of pop-cultural consciousness contained within AI.

Until then, I’ll keep hunting him, sharpening my harpoon-like prompts.

Coda: Using Dreambooth to Train Stable Diffusion on Steamboat Mickey

While writing this piece, it occurred to me that the best way to generate accurate Steamboat Mickeys was two-pronged: to custom train an AI on images while conversely never calling on Mickey by name in the prompt.

This has now led to some of my better results, and I’m convinced that the solution lies in this direction. As Captain Ahab might say, “Thar she blows!”

Jim the AI Whisperer (2024)
Jim the AI Whisperer (2024)

The trick was to make a collection of screen-grabs of the Steamboat version of Mickey, then use Dreambooth to put them into Stable Diffusion, much as you might any person or object you’d like to consistently recreate in AI.

I removed all the vowels (‘Stmbtmcky’) to make a unique identifier, and so it wasn’t tempted to draw on existing training of images of Mickey Mouse.

So now I have a custom AI model of Steamboat Mickey. However, there is some overfitting, as we don’t have much footage and it’s all very similar. And I couldn’t rely on prior preservation by class as it risked evoking the monolith that is Mickey. Future steps would be to add Plane Crazy Mickey for variation. Another option might be to try textual inversion and inject a new word (‘Stmbtmcky’) to fine-tune the textual embedding of the model.

I’m going to keep at it. Ultimately, I’d like to create an AI animated short sequel to Steamboat Willie. It’s time consuming, and quite costly, so if you’d like to be a patron of this project, I’d be delighted to be fuelled by your support. Visit my Buy Me a Coffee page and drop a coin or two ☕️ 😉Your support keeps my brain buzzing with new strategies and solutions.

You can also show your support by Clapping and Sharing this article. Remember, if you press and hold Clap, you can applaud for a writer! There’s no limit, so go applaud all your favorite writers on Medium.

You might enjoy these Disney AI articles from Jim the AI Whisperer:

Artificial Intelligence
Design
Technology
Machine Learning
Art
Recommended from ReadMedium