In order to create the images used in this project,
I first began by selecting random images from the LAION-AESTHETICS Database
that had an aesthetics score of 6 or higher. Next,
these images were processed through a Clip Interrogator using the
ViT-H model in order to create the prompts presented
for each image. Finally, the prompts were passed
onto the Stable Diffusion 2.1 Model
in order to generate their respective images. Each
prompt was passed through 5, 25, 50, 100, 500 and
1000 inference steps with a guidance scale of 7.5 to
generate images with self-assigned difficulty scores
of 1, 2, 3, 4, 5, 6 respectively. The work was done
on a 5950x CPU and 3090 GPU with 128 GB of RAM
available.
The LAION-AESTHETICS images were selected because
they're apart of the database used to train Stable
Diffusion so it made sense to provide prompts
similar to what Stable Diffusion had seen before in
order to create the most impressive images. From
there, clip-interrogator was chosen as the prompt
generation model in order to create descriptive
prompts tailored to the image generation models
available. Lastly, Stable Diffusion was chosen over
DALL·E 2 because of its realistic look tailored to
some of the non-painting images chosen. After
comparing many of the images generated, DALL·E 2
excelled at creating paintings while Stable
Diffusion was subjectively better at generating
photographs.
AI art has been a hot topic in the online world as many AI tools are being shared in various different communities. One of the most talked about topics is currently AI art and how it challenges what the world perceives as art. There have been many debates over the legitimacy and ethicality of AI art since many of the AI models used to create these images were trained on art and images that were scarped off of the internet without the permission of the owners. As a result, artists are defying the creators of these models because they believe their intellectual property was stolen in order to generate these images.
There are various articles that you can real in order to better inform yourself on this issue. Here are a few that you can read:
- NYT: Are A.I.-Generated Pictures Art?
- NYT: An A.I.-Generated Picture Won an Art Prize. Artists Aren’t Happy.
- Kotaku: AI Creating 'Art' Is An Ethical And Copyright Nightmare
- Kotaku: Adobe Embraces AI Art, Can Get In The Bin
- Tech Crunch: AI art apps are cluttering the App Store’s Top Charts following Lensa AI’s success
The goal for this project is to showcase what AI is doing, both how great it is but also the many complexities especially in regards to the ethicality of it. I’ve been working on AI/ML/DL projects since my freshman year at Davidson and it's what I see myself doing in the future so this is a topic I really enjoy. There are many great applications for AI that are transforming the world even though we might not see them but there are also many downsides to it as presented in the readings and short blurbs above. I am a part of an online community filled with artists so when the topic of AI came up, it really interested me that many of them saw AI in such a negative way so I began exploring the topic a little more. I was then linked to some articles and saw how AI was now generating images and starting to destroy some of the smaller artists. A few months go by and the topic is now more relevant than ever as the Lensa App on iOS became such a big hit at the end of November that it reached the #1 photography app on the app store in a matter of days. This app brought along with itself much criticism due to it being an AI app that transformed existing images into AI art-style renditions. Next, in the beginning of December, ChatGPT took the internet by storm, as a record breaking 1 million users tried the AI text generating model in under a week. ChatGPT is a great conversation and text generation algorithm that can provide excellent responses to some prompts but it also has many downfalls that many people don't see. Through this project, I wanted to showcase the downfalls and negatives of AI tools so I decided to focus on the generation of AI images that come from models trained on images scraped from the internet. Many of those images come from smaller artists that had no say on the use of their images for these models. Through the comparison page and data page, people are able to see how powerful image generation models have become but they can also see the many mistakes that those models make. Through the about page, I hope to be able to inform people on the topic so everyone can understand the complexities of AI models and artists.