Amy Anderson
West Chester University
Inspired by the interplay between words and images in medieval illuminated manuscripts, this assignment invites undergraduate students to work with an image-generating LLM to create their own illuminated manuscript of a passage from a course reading. By generating images that comment on and subvert the course reading, students are asked to explore the inventional relationship between words and images in both LLM prompts and multimodal compositions. Although the assignment was designed for a Medieval Women’s Culture course, it could be adapted to any rhetoric or composition course that considers multimodality.
Learning Goals
Original Assignment Context: Final assignment in an upper-level elective English department Medieval Women’s Culture course, covering women in medieval rhetoric and literature
Materials Needed: Accessible image-generating LLMs for students to use (ChatGPT, Midjourney, Canva, etc.)
Time Frame: 3-4 weeks
Overview
After spending most of the semester considering how manuscript culture affected the transmission and reception of medieval texts, the sixteen students in my Spring 2024 Medieval Women’s Culture course spent the last three weeks of the class learning about and creating digital illuminated manuscripts of passages from our course readings, using image-generating LLMs to help with the illuminations. As Michael Camille argues in “Making Margins,” the first chapter of Image on the Edge, manuscript illuminations influence the reader’s perception of a text by illustrating, expanding, critiquing, and subverting the words – as well as scandalizing the reader and making use of material imperfections in the page. Creating illuminated manuscripts is thus an ideal way for students to showcase their interpretive skills and to think critically about the relationship between words and images in multimodal compositions.
My motivation for asking students to use image-generating LLMs to design their manuscript illuminations was two-fold. On a practical level, using LLMs made the assignment more accessible for everyone, including students who were not artistically inclined. Experimenting with LLMs also enriched the assignment by encouraging further critical thought about the word-image relationships. When students engineer word-based prompts, hoping that an AI will generate a particular image, the results are not always what they expect, leading the students to reconsider the language of the prompt.
At the outset, students were hesitant to approach the assignment because, while they had plenty of experience using LLMs to generate words, no one had used it to generate images. As we moved through the unit, the students regularly shared the strategies they were developing for working with the LLMs. Some found that reprompting got them closer to their desired image (“Make the man younger,” Make the border .5 inches narrower,” etc.). Other students found that the unexpected details in a generated image (such as a nun’s disturbingly creepy smile) could serendipitously accentuate their interpretation of the text. Yet others were dismayed by persistent incongruent details, such as extra limbs or tails on humans. Some students even opted to augment AI-generated images with hand-drawn details when they couldn’t get exactly what they wanted.
Based on feedback at the end of the semester, the project was a success. As a whole, the students found the process of working with an image-generating LLM thought-provoking, particularly because it revealed LLMs’ current capabilities and limitations. They enjoyed the chance to be creative and experiment, while still thinking critically about the class readings. Some students expressed concern about the cost of working with ChatGPT and Midjourney, the two LLMs that I recommended they try. A few had found a free work-around and used the generative AI embedded in Canva. While this solved the affordability issue, Canva’s AI currently limits the number of possible outputs and thus didn’t let them experiment as much as they would have liked. The assignment has only been taught once, and a possible future solution would be to give students the option to work in groups to share costs.
At the beginning of the unit, the class read and discussed Michael Camille’s “Making Margins,” the first chapter in Image on the Edge, which surveys the various ways that medieval manuscript illuminations interact with the words in the manuscript. The chapter includes lots of examples, so students gained a framework for what medieval illuminated manuscripts look like and how they function. The students were then given the following prompt:
Now that we’ve read the chapter from Michael Camille’s Image on the Edge and looked at various example manuscripts, it’s your turn to practice the enlightening and subversive medieval art of manuscript illumination. From our course readings (Beowulf, The Art of Courtly Love, Julian of Norwich’s Divine Revelations, The Book of Margery Kempe, Canterbury Tales, etc.) choose either one passage of 400-500 words or two passages of 200-250 words each, and create two illuminated manuscript pages. The font and layout of your words should be meaningful, and each page should include at least four illuminations that offer commentary on the text and suggest how readers should interpret it. At least two of your illuminations should subvert the text, and your illuminations should be created using ChatGPT, Midjourney, or another image-generating LLM. The manuscript should be submitted as a Word document or pdf.
You’ll hand in your illuminated manuscript with a 700-900-word explanation of how your font and illuminations both comment on and offer interpretive suggestions for the central text(s), as well as your process of generating the illuminations. During the last week of the semester, you’ll give a short (5-10-minute) presentation discussing your manuscript and your experience working with generative LLMs; your presentation will count for 20% of the project’s final grade.
Once students had chosen the passages that they wanted to illuminate, we spent a class period discussing how LLMs like ChatGPT and Midjourney are trained and generate responses to prompts. I showed students the sometimes dubious results of my own attempts to use ChatGPT to generate manuscript-style images of medieval monsters, as well as the process of reprompting to refine images. We used my examples to discuss prompt engineering and LLMs’ capacity for working with images. As students were designing their manuscripts and writing their reflections, we spent part of every class troubleshooting and sharing strategies for working with the LLMs. After a round of small-group peer review, students ended the semester presenting their work and celebrating each other’s creativity.
To complete the assignment, students need access to a word processing program and an image-generating LLM, such as ChatGPT or Midjourney. This assignment could be easily adapted to any course that deals with multimodality and has a learning goal of exploring the inventional relationship between words and images. Illuminating nearly any course reading in any genre will invite students to consider how images might shape the interpretation of words, and vice versa. Experimenting with word-based LLM prompts also invites students to consider how words can describe and evoke images, as well as how the evoked images can reflect unconscious assumptions and biases. The Camille reading was appropriate for a class focused on medieval texts, but the principles from Camille’s chapter could be adapted for classes with different emphases.
Camille, Michael. Image on the Edge: The Margins of Medieval Art. Reaktion Books, 2004.