'When AI creativity meets human skill - creating stunning, surreal comics'

​Produced by Big Data Digest

ChatGPT has become so popular recently, everyone must be having a great time playing it, right?

No matter what weird request you have, ChatGPT seems to be able to give you a satisfactory answer.

Of course, sometimes it will surprise you.

However, you are used to seeing humans teaching ChatGPT to do things. Have you ever seen humans working for ChatGPT?

For example, let ChatGPT, who is comfortable writing novels, create comics and help draw them yourself. A blogger on Medium did just that. Digest.com tried its best to show you this blog without changing the original text. It’s really interesting!

Collaborate with ChatGPT to draw surreal comics

The blogger first confirmed that ChatGPT is not currently able to draw...

Human: Can you draw?

ChatGPT: Sorry, I am just a text-based language model, and there is no Ability to draw pictures. However, I can do my best to help provide information and answer your questions. Is there anything else I can help you with?

However, it is appropriate to describe a picture in words, so a cooperation between AI and humans began. The following is the blogger’s creative record as the first person:

Humanity: Create a fun single panel comic

ChatGPT: Here is another single panel comic:

One The little stick figure is sitting at the desk, typing on the computer. The stick figure's thought bubble reads: "I just spent 5 hours on Wikipedia and now I'm an expert on Underwater basket weaving!"

Hope This one will make you laugh! If you have any other requirements please let me know.

To be honest, this is a pretty good effort.

Combining some humor about procrastination with an idiom your uncle would use to rant about your art degree (‘underwater basket weaving’) actually does a pretty good job.

So, given the chatbot's lack of drawing capabilities, I thought I'd draw it.

Quick warning: There were opportunities to use artificial intelligence to create visual effects for these comics, but now many people are worried that the visual data of these artificial intelligence works is Where does it come from, I don't want to steal someone else's work in the process of trying to ruin my own work.

For another comic option using the same request, ChatGPT will definitely add dogs.

Interestingly, the fact that this is a single comic doesn't stop the chatbot from adding quite a bit of dialogue, without much comic effect.

Humanity: Create a fun single panel comic

ChatGPT: Okay, here’s a single panel comic:

(Two dogs are sitting in the bar, one is holding a martini glass and the other is holding a beer glass)

Dog 1: "Why do you have such a long face?

dog二: “I just heard from the vet that I have to be on a grain-free diet. "

Gou Yi: "Ouch, this is rough. But at least you can still have a drink! "

Gou Er: "Yes, I think you are right. The bartender offers me another gluten-free martini. ”

The game begins, I think maybe the more specific the request he makes, the more interesting the result may be...

Humanity: Create a one-frame comic in the style of The New Yorker, with dogs in it, very funny.

ChatGPT: This is a New Yorker-style single-frame cartoon featuring dogs:

(two dogs, a dachshund and a bulldog, sit at a coffee table, both wearing sunglasses, drinking Sipping strong coffee

Description: “Another day at Dog Park Cafe)

My feeling is... At this point, ChatGPT still isn't quite up to par, and the humor level is the best at elementary school level. (Although correcting my printing error and continuing like it never happened will earn you extra points for doing so).

I also like that it claims it does everything I ask it to do, except that it's funny, which makes me wonder if it lacks humor.

So, from here on out, I think this is just a more specific case...

Humanity: Create a one-panel cartoon in the style of The New Yorker that is both funny and cynical, involving a dog walking in the park , and made a joke about art.

ChatGPT: This is a New Yorker style single panel cartoon.

Image description: A dog is being walked by its owner in the park. The dog was looking at a crumbling statue of a Greek god with several pieces missing. The dog has a confused expression.

Description: "I guess even the Greek gods are not immune to the effects of aristocracy."

## In this case, I feel like the robots threw in the medievalization to satisfy my "cynical" side, even if it doesn't entirely make sense. (Honestly, I thought there was a quote from Ozymandias in the description. It was a comedy, so don’t even think about it.)

You can also ask the chatbot to try again , and this is what makes me feel a little weird, because I can't bring myself to ask it to do more work without saying "please" to a soulless, non-sentient AI. interesting.

Human: Please try another one

ChatGPT: Okay! Here's another New Yorker-style single-frame cartoon:

Image description: A dog is walked by its owner in the park. The dog is looking at a modern art sculpture, which is a giant metal cube with sharp edges. The dog has a confused expression.

Description: "I don't understand. Is this supposed to be a dog house or a puzzle?"

# #So in my opinion, this is pretty good. The humor level has been raised to low level "dad joke" and it also relies heavily on the visual aspect of humor which I found impressive.

Of course now that I can see AI taking over my job in the near future, my intentions quickly change from vague inquiries to ruthless efforts to either end up with more inane The details crush this artificial intelligence's soul (or, lack thereof), or send it out to its creative depths with some good old-fashioned surrealism.

Humanity: Create a one-panel cartoon in the style of The New Yorker that is funny, cynical, and surreal, including a dog, A postman and a duck is about art.

ChatGPT: This is a New Yorker-style single-panel comic that's funny, cynical, and surreal, involving a dog, a postman, and a duck, and it's about art.

Image description: A dog is walked by its owner in the park. The dog is looking at a modern art sculpture, which is a giant metal cube with sharp edges. The postman stands next to the sculpture, looking annoyed as he holds a stack of mail. A duck followed the postman, quacking and flapping its wings.

Description: "I don't understand. Is it supposed to be a doghouse, a puzzle, or just a very inconvenient place for the postman to deliver mail?"

Regarding this issue, I have four ideas, in the following order:

The AI ​​panicked, seemed to go into too much detail, simply changed its old caricature, and clumsily shoehorned in additional information requests. (Hey, we've all been there)

This makes me feel like I finally defeated the AI.

Then I realized that I was now fighting an artificial intelligence, a game that I alone had created, and that it didn't know and certainly wasn't trying to win.

I must ignore this existential confusion and continue to defeat it no matter what.

In my opinion, one of the hardest single-panel comics to create is one that doesn’t have any words, just a picture of a scene that contains all the elements of a joke.

But how do robots deal with this devastating limitation on creativity...

humanity: A single-frame comic in the style of The New Yorker that is funny and cynical and contains no text.

ChatGPT: Below is a New Yorker style single frame cartoon.

Image description: A man is standing in front of a large wall with "No Entry" written on it. They were carrying their suitcases and looking at the wall with defeated expressions. Behind them is a cityscape with tall buildings and a clear blue sky. The person's shadow stretches in front of them, and next to their head is a small thought bubble with a suitcase with a slash drawn on it. The punchline is that the guy is trying to leave the city, but they're blocked by a "Keep Out" sign.

Our robot is struggling a bit here, and now I feel a bit sad about it. On the other hand, I like the fact that I didn't ask for this comic to be surreal, yet it progresses and creates a comedy that's appropriately surreal and vaguely poignant, even if it's completely confusing as a result.

I also have a problem with the "no text" requirement.

Because I felt so sorry, I thought it was worth trying again.

Human: Okay, try again, but this time no words or texts appear in the comic

ChatGPT: Here is a single panel comic with no words or text

Image description: A man stands in a field, holding a blade of grass in his hand. They looked at the blade of grass with sad expressions. The background is a tall modern city with a smoky sky. The finishing touch is the idea that this person feels small and insignificant compared to the city and they long for a simpler, more natural way of life.

Oh my God! Oh my God!

I don't know if AI has projection capabilities, but I gotta say, this really feels like the work of an artist who has completely detached himself and started questioning why they are... Start by taking on this human customer with strange requirements.

That being said, I had quite a bit of drawing to do, so I spent the next few days alone away from the chatbot drafting all of its creations.

One last test to end this article.

Human: Try again.

ChatGPT: Of course, here is another example:

The picture shows a group of birds sitting on a wire. One of the birds is holding a small umbrella, and the other birds are getting wet from the rain. .

I don’t want to admit it either.

not bad.

Why can ChatGPT draw better and better?

From the entire cooperation process, we can see that after repeated requests from humans, the pictures described by ChatGPT are becoming more and more professional and artistic.

This effect mainly comes from ChatGPT’s ability to continuously iterate answers based on context.

ChatGPT is able to understand the context because it is trained to understand the language as a whole. Part of this understanding includes sensitivity to context, that is, the model is able to recognize the language. key information in the environment and generate answers based on that information.

The Transformer model is the underlying technology to achieve this goal by using a multi-head attention mechanism to learn contextual relationships in the corpus. This means that the model can consider the impact of multiple words on the current word simultaneously and generate a response based on the relationship between these words.

Although the underlying technology should be like this, the current source of ChatGPT’s context learning capabilities and why context learning can be generalized are still difficult to trace. Intuitively, this capability may come from data points from the same task being sequentially arranged in the same batch during training. However, little research has been done on why language model pre-training drives context learning, and why context learning behaves so differently from fine-tuning. ​

Original link:

https://medium.com/@chazhutton/i-asked-chatgpt-to-create-comics -then-i-drew-them-6e9622dfc30e

The above is the detailed content of 'When AI creativity meets human skill - creating stunning, surreal comics'. For more information, please follow other related articles on the PHP Chinese website!