Inspiration
With the advent of AI generated images in recent months, we were very intrigued by their potential. We've noticed a few captcha images featuring AI generated images. However, we were not quite sure what made these images so blatantly AI generated. Sometimes it was because they featured something impossible, like a rabbit swimming in the ocean. However, other AI generated images were extremely beautiful. Back in August, Jason Allen, a digital artist, submitted a work he created using a artificially generated image to the Colorado state fair. This caused a lot of controversy online surrounding the legitimacy of his submission. What makes art, real art? our goal is to explore the line between real and fake. Why were some images more convincing that others? What can we do to make more AI generated images convincing?
What it does
This project primarily focuses on images generated by Stable Diffusion, a service which creates AI generated images. Our project pulls images from Lexica, an API which features images generated using Stable Diffusion and their prompts. It also pulls images from Laion, which features images used to train Stable Diffusion. Then, it presents these options to the user and asks that they pick which one they found the most realistic. This information is stored in a database, which we can then use to generate more convincing AI generated images.
How we built it
We used Django to create the web interface, and wrote a python script in order to access the Lexica and Laion Datasets. Lexica API allows us to access AI generated images. The Laion datasets allows us to access real images that were used to train Stable Dissusion. We stored the image data and results sin an SQLite database. We combined these two features together in order to collect data on the AI's convincingness.
Challenges we ran into
First, we did not now how to use Django, so we had to take a tutorial on how to create a website and database using it. The information from the API's also needed to be reformatted, since some of the information was not in ASCII format.
Accomplishments that we're proud of
We learned how to use Django, and reference APIs. We also came up with an idea which is incredibly useful for digital artists and those interested in AI.
What we learned
We learned how to create a website using Django, how to reference APIs in this framework. We also learned how to create convincing AI generated images.
What's next for Which image is real?
Next, we would like to create an algorithm to look at the prompts used to create the most convincing AI generated images, and integrate those into the program. This will result in more and more realistic prompts, which will eventually be able to create more realistic images. We would also like to polish the interface more, so it will be more accessible to users. Another goal is to create a data visualization for the results we have collected. Lastly we would like to deploy this website on the internet.
Log in or sign up for Devpost to join the conversation.