Building The Legend With AI
The AI revolution is upon us, for good or for ill.
We at Legendao have tinkered with several AI tools and are looking forward to possibly incorporating some of them into our creative process. We wanted to give you a rundown of what we learned and what we’ll use.
But we’re also going to give a bit of background in how Midjourney and other AI art generators manufacture their product, IP concerns and how implementing AI without understanding its downside could do more harm than good.
AI: Boundless Creativity or Glorified Regurgitation?
The following question was put to ChatGPT.
Anyone with a tinge of critical thinking ability can already see why ChatGPT did not exactly bring up a goldmine of thought leadership with its reply. That response was about as mediocre as you can get, but let’s stick with Midjourney for now.
ChatGPT describes it as thus (with lots of editing on the author’s part because of its bad grammar and repetitive phrasing):
“Midjourney uses a technique called style transfer to generate its art. The specific mathematical formula used in style transfer is a type of neural network called a convolutional neural network (CNN). It works by analyzing the input image at different scales (or “layers”) and then uses that analysis to make predictions about the image.
“The algorithm involves training a CNN on a large dataset of images. The CNN is trained to recognize the underlying style and composition of the images in the dataset. Once the CNN is trained, it can be used to analyze a new image and “transfer” the style of the images in the dataset to the new image.”
The latter output was generated on Friday, January 13.
As of Sunday, January 15th, however, try to type Midjourney into ChatGPT and it will deny any knowledge of such a program existing:
Despite this, ChatGPT describes the process with which AI art generators pull together their product as follows:
The process may involve several steps such as:
Data collection: The program may collect image and text data from various sources, such as online databases or user inputs.
Data pre-processing: The collected data may be pre-processed to remove any irrelevant or redundant information and make it ready for input to the AI algorithm
Training the algorithm: The algorithm may be trained using the collected data, adjusting the weights and biases of the neurons in the network using backpropagation and gradient descent algorithms.
Generating new art: Once the algorithm is trained, it can generate new images based on the input data. The algorithm may use techniques such as GANs, neural style transfer, or deep learning to generate new images.
Output: The generated image is output to the user.
“It’s important to note that different AI art programs may have different specific methodologies and architecture. Therefore, this explanation is a general overview of what might happen when you enter an /imagine prompt in an AI art program.” — ChatGPT
Obviously, there are billions of images online that possess proprietary IP. This could mean legal implications for anyone creating AI art for commercial purposes.
The first class-action lawsuit has already been filed.
“We’ve filed a class-action lawsuit against Stability AI, DeviantArt, and Midjourney for their use of Stable Diffusion, a 21st-century collage tool that remixes the copyrighted works of millions of artists whose work was used as training data.” — Matthew Butterick
Navigating the AI/IP Minefield
With these facts in mind, it’s easy to grasp why there’s no shortage of IRL artists coming out against AI art.
ArtStation has been taken over by protesting artists who post the same image over and over again: a large red “no” sign covering the word “AI” paired with a caption that reads “NO TO AI GENERATED IMAGES.”
Read more about that on Vice.
It doesn’t seem like OpenAI has a leg to stand on.
I mean, except the 10 billion dollars Microsoft is about to pump into them, but that’s something else.
Ditching the Dirt, Keeping the Gold
All this notwithstanding, any creative business that ignores the capacity for AI to streamline workflow is going to be left behind.
There are amazing things AI can do for any creative enterprise. Tools like pitch deck creators, AI video producers, Dungeons and Dragons dungeon masters, and anything else you can think of are all the rage. People use them because they’re efficient (though as you’ve seen above, they churn out varying levels of quality).
But there’s only one Legendao is seriously considering usage of at this time.
We’re concerned about using AI for game assets because of two reasons. We don’t want to rip off artists because of idealistic concerns. We also don’t want to face any repercussions. At this point, basing a commercial enterprise on AI assets portends legal disaster. The implications for the creative community are utterly mind-boggling. And that’s a minefield we don’t want to enter.
That’s why we’re:
- using our own IRL artist to create the assets actually used in the game
- having our community vote on human-written threads to decide our storyline
- rewarding our community for submitting their own quests rather than depending on some AI service to make them up
Legendao Universe is the game that keeps on giving. It is an ongoing project whose game will be fortified by the creative potential of its community.
Genie and other AI programs we are looking at like Scenario help us keep it real because they allow us to only use our own images and data to build assets from.
We retain our creative integrity, honor the IP rights of others and authentically engage our community in a way that builds their bags and gives them a good time.
So are our Yetis, by the way. But that’s for another blog post.
Stay tuned because we’re about to drop another post with all the gameplay details.