Inspiration
There are many AI artists that I have become a fan of. Each artist is unique, but they share a common trend in how they use the weirdness of AI art to their advantage. Rather than using AI art to replicate something that already exists, they lean into the medium as something completely new, and are able to continuously produce unique results. I firmly believe that is the best possible use of image generation, and wanted to recreate this embrace of the strange in my project.
What it does
This app goes through four stages of prompted AI widgets, with each widget building on what came before it. First the user chats with a mad scientist character who is "studying" them. They are asked five very strange hypothetical scenarios, and the scientist makes observations dependent on the user's answers. The next widget will create a bulleted list of all animal comparisons made by the scientist. Next a new widget generates an image prompt of a hybrid-animal made up of the listed animals. Finally, a realistic image is created to showcase the new sci-fi inspired creature that the mad scientist has made to represent the user.
How I built it
This project is a remix of the community-made chatRPG. All of the original widgets (chat, summarize, image prompt, and image generate) are all still in play and interpreting information from the previous widget. The prompt for each widget was drastically changed, but still follows the same pattern of chat interaction, breaking down elements of the interaction, using the elements to create an image, and then generating said image. First, the base idea of generating a hybrid animal based on the user's personality was implemented in a very basic way. The next step was to optimize the chat and summary widgets to work together. The chat prompt asks for a very specific format of mentioning an animal, which the summary is prompted to recognize. The summary then generates the animal list and reduces mention of any unnecessary details, to prevent confusing in the image prompter.
Challenges I ran into
It is no secret that many generative AI models are not very keen of the horror genre. My goal was to make the hybrid-animals slightly uncanny and off-putting, but many of those phrases get rejected by the LLM, preventing an image prompt from being made and resulting in a completely random image being generated. I found success by focusing on phrases like "sci-fi" and "alien" to get more unique results. Unfortunately, it was at this point that the AI began to have a problem with creating a new species. It was a mix of questionable morals and the possibility of spreading fake news with the realistic format. This was remedied by continuously emphasizing that this animal is for a fictional story. Additionally, I had to move the "photo-realistic" portion of the prompt into the image generation widget, effectively separating the mentions of realism and fiction to prevent the prompt being seen as false information.
Accomplishments that I'm proud of
I find the mad scientist's way of communication to be very charming and human-like, with specific anxieties and quirks embedded in its behavior. Additionally, the hypothetical scenarios presented were done with a basic prompt of "x in y scenario" requesting varying characters and situations to fill in x and y respectively. Not much was needed to modify this prompt, as it seemed to immediately create very random questions that I found to fit in well with the rest of my project. And most of all I am proud of the images that have been generated. They are able to walk the line between sci-fi and horror that was so hard to achieve, but I am absolutely in love with the results.
What's next for Mad Science Zoo
If I am able to create a very specific form of a hybrid animal that exists in latent space, I could go on to generate the same animal in specific scenarios, or with slight tweaks added to it.
Built With
- generativeai
- partyrock
- stablediffusion
Log in or sign up for Devpost to join the conversation.