Prompt Exploration of AI Image Diffusion Models for Students

For several years now, I have had my students explore AI image generation using text prompts. It’s been fascinating to see how the technology has progressed and what students can create with it. Before the technology became so prominent, I used to have the students do a small competition using Runway.

At first, it was difficult for students to get good results, but with technological improvements, students were able to create increasingly impressive images. Perhaps more importantly, the initial “wow” factor has diminished. Around 2021-2022, almost no student had tried an image generator. Now, although they tend to use large-language models like ChatGPT more, they are generally less amazed by the output.

For those who are unaware, numerous models for image generation give slightly different results based on the same prompt. When working with students, it’s important to discuss the ethical implications of AI-generated images, such as potential biases in the training data. Often, they believe these biases only occur with images of minorities, but the results affect all images. Consequently, I found one of the best ways to get students to understand the limitations is to have them generate images using the same prompt with two different models. We then create a slider showcasing the different results.

I realized I hadn’t taken the time to systematically explore these models myself, so I decided to try out different diffusion models. I found that StableCog seemed to provide a generous free tier with various diffusion models at the time, and sliders comparing outputs can be created fairly easily in Flourish or Juxtapose. I especially wanted to try vector graphics since those usually require a lot of manual work in programs like Adobe Illustrator.

None of these models advertise themselves as being for vector graphics, but it is worth exploring to see how well they perform, and students seem to enjoy pushing these models to their limits. Anyway, here are some differences between the models using the same prompt. I drew from StableCog’s Gallery for the prompts but added terms related to vector graphs: