{"id":12633,"date":"2022-05-04T19:00:42","date_gmt":"2022-05-04T17:00:42","guid":{"rendered":"https:\/\/www.sciencefocus.com\/?p=118782"},"modified":"2022-05-04T19:14:11","modified_gmt":"2022-05-04T17:14:11","slug":"we-badly-described-cartoon-characters-to-an-ai-heres-what-it-drew","status":"publish","type":"rss_feed","link":"https:\/\/c01.purpledshub.com\/bbcsciencefocus\/rss_feed\/we-badly-described-cartoon-characters-to-an-ai-heres-what-it-drew\/","title":{"rendered":"We badly described cartoon characters to an AI. Here\u2019s what it drew"},"content":{"rendered":"

<\/p>

By Alex Hughes\n \t\t<\/p>

Published: Wednesday, 04 May 2022 at 12:00 am<\/p>

\n\n

The use of artificial intelligence<\/a> (AI) has made its way into just about every industry in the world, solving complex problems, learning difficult skills in seconds, and generally blowing the minds of humans with advanced calculations we could never dream of.<\/p>\n

But while most people are looking to use AI for the good of humanity, bettering society and solving the big questions of today, I decided to use it for a higher purpose \u2013 recreating some of the most recognisable cartoon characters through an image-generation AI.<\/p>\n

How it works\u00a0\u00a0<\/strong><\/h2>\n

Back in January 2021, the very intelligent people over at OpenAI created a program that, when fed a string of text, could use this to generate an image. This could be \u2018a lion on a sofa\u2019, \u2018 a black hole in a box\u2019, or some other equally weird prompt. However, this technology would often produce blurry images, or struggle to fully understand the prompt that it was given. In April 2022, OpenAI launched the second iteration of this product \u2013 Dall-E 2<\/a>.<\/p>\n

Now, you are able to get high-quality images based on your worded prompts, offering highly detailed images in a matter of seconds.<\/p>\n

The important thing to note with this technology is that, the more information you give it, the better the end result will be. However, I lack the creative skills to really paint a verbal picture\u2026 and that is pretty noticeable with my AI-generated cartoon characters. The software is able to perform at its best when it is given a style of art, and plenty of description. I instead opted for vague art terms and a lack of detail \u2013 what could go wrong!<\/p>\n

BoJack Horseman<\/strong><\/h2>\n

Prompt:<\/strong> Horse in a suit as a oil painting<\/em><\/p>\n

\"""\"<\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/picture><\/div>
(Left) Dall-E 2\u2019s image of a horse (right) Bojack Horseman<\/em> \u00a9 OpenAI\/Netflix<\/span><\/figcaption><\/div>\n

Let\u2019s be honest, based on the description \u2018Horse in a suit\u2019, Dall-E 2 nailed my description \u2013 that is indeed a horse in a suit, and my request of an oil painting style was nailed absolutely perfectly. I would even like to think that if there was a season where BoJack Horseman<\/em> was a mid-century nobleman, my attempt at getting an AI to recreate him would be absolutely bang on.<\/p>\n

I think if Dall-E 2 had been given more information on this one, we could have got a perfect recreation of BoJack in an oil painting!<\/p>\n

\"""\"<\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/picture><\/div>
Dall-E 2\u2019s other attempts at creating horses were just as accurate \u00a9 OpenAI<\/span><\/figcaption><\/div>\n

There were two other examples of horses in suits, both also in the style of an oil painting. I think we can all agree that they are just as dapper as the AI\u2019s other attempt \u2013 who knew horses could style a suit so well?<\/p>\n

Teenage Mutant Ninja Turtle<\/strong><\/h2>\n

Prompt:<\/strong> Turtle wearing a black mask and holding a sword and pizza as a cartoon<\/em><\/p>\n

\"""\"<\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/picture><\/div>
(Left) Dall-E\u2019s rendition of a Teenage Mutant Ninja Turtle<\/em> (right) the original \u00a9 OpenAI\/Paramount Studios<\/span><\/figcaption><\/div>\n

It worked\u2026! Well, it is at least recognisable this time. The prompt of \u2018a cartoon\u2019 did a lot of the heavy lifting here, generating a more child-like version of the original Teenage Mutant Ninja Turtles<\/em>.<\/p>\n

By giving it a bit more information than I did with the horse, Dall-E was able to create a closer attempt.<\/p>\n

While this is by no means an exact replica, it\u2019s another impressive attempt based on a limited prompt.<\/p>\n

Donald Duck<\/strong><\/h2>\n

Prompt: <\/strong>Duck with blue shirt and red bowtie as a pencil drawing<\/em><\/p>\n

\"""\"<\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/picture><\/div>
(Left) Dall-E 2\u2019s depiction of a duck (right) Donald Duck<\/em> \u00a9 OpenAI\/Disney<\/span><\/figcaption><\/div>\n

What I should have learned from the two above attempts is that the cartoon art style was the way to go here\u2026 I did not learn. This is a blessing and a curse because, while this is easily the least accurate attempt, it is also my favourite.<\/p>\n

These are both ducks, they both have red bow-ties and blue shirts, they are both even done in different styles of drawing, but they could not look any less alike.<\/p>\n

While I\u2019m not sure how much extra information I could have fed (other than a hat), the end image seems like an obvious result \u2013 it is exactly what I asked for after all.<\/p>\n

\"""\"<\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/picture><\/div>
Dall-E 2\u2019s other attempts at a duck \u00a9 OpenAI<\/span><\/figcaption><\/div>\n

Yogi bear<\/strong><\/h2>\n

Prompt: <\/strong>Bear with a green hat and green tie as an animation<\/em><\/p>\n

\"""\"<\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/source><\/picture><\/div>
(Left) Dall E 2\u2019s attempt at Yogi Bear<\/em> (Right) the real Yogi Bear \u00a9 OpenAI\/Warner Bros<\/span><\/figcaption><\/div>\n

Okay, we\u2019ve got really lost on this one. In an attempt to find a Yogi Bear<\/em> doppelg\u00e4nger, I seem to have ended up with a St Patrick\u2019s Day mascot.<\/p>\n

Once again, all of the points are correct. They are both bears with green ties and hats, they just look nothing alike.<\/p>\n

At least I was more on brand with the art style this time!<\/p>\n

Did Dall-E 2 work?<\/h2>\n

While I think it is safe to say this was a failed mission, I don\u2019t think it has anything to do with the ability of Dall-E 2. The software has proved its impressive abilities across a host of prompts, art styles and situations, the lack of similarity comes down to a few key issues.<\/p>\n

Firstly, I went into this thinking Dall-E would be on the same wavelength as me. As someone looking to recreate cartoon characters, the prompts are obvious to me \u2013 of course \u2018a duck with a blue shirt and red bowtie\u2019 is Donald Duck! But to a program that is very literal in nature, I am asking for something completely different.<\/p>\n

Artists with a much better imagination than me have also shown that, with a more detailed prompt and understanding of the platform, you can get some much more impressive results.<\/p>\n

Dall-E 2 also has a feature that would have given me a much more accurate result. When you insert an image, it will create hundreds of versions of it in its own style. In retrospect, a much more logical approach to this task.<\/p>\n

Read more:<\/strong><\/p>\n