Prompt design, a critical component of text-to-image synthesis, involves crafting text prompts that elicit visual responses from image generators. There is no universal formula for designing effective text prompts: a prompt might be designed through rapid, high-volume experimentation, or via calculated approaches. Nonetheless, the effectiveness of a text prompt is largely determined by the chosen model's¹ text-image alignment capability, as measured by benchmarks such as compositionality², factuality³, spatiality⁴, cardinality⁵, etc.

Assessing a model’s capabilities and limitations is a crucial step in determining its appropriateness for classroom instruction. What are the goals of your planned lesson? What concepts might students visualize to meet your stated goals? Does the chosen tool generate the subject accurately, semi-accurately or not at all? Is the tool designed with appropriate guardrails⁶?

In this section, we propose a sustainable, process-oriented approach to prompt design. Our worksheets⁷ will guide you through each stage of the design lifecycle. All worksheets are linked to buttons on the bottom left of each section. Click on the button to download respective worksheets.

DOWNLOAD THE K-12 SUITABILITY WORKSHEET

To design an effective text prompt, students will identify core elements linked to your learning objectives:

1. Subject(s): What is the primary subject(s) of the lesson? For example: If the lesson is about the Harlem Renaissance, the subject might be a geographical location in Harlem. Perhaps, The Apollo Theater?
   
   Current models are better at object-rendering than text-rendering, and handle material/color/size attributes better than count/shape attributes.


2. Scenario: Where is the subject situated? What time period?
   
   TEXT PROMPT: Apollo Theater in Harlem, 1930s


3. Attribute(s): Once the core elements are defined, students should qualify or add specificity to the subject, and/or it's context.
   
   TEXT PROMPT: Apollo Theater in Harlem wrapped in Danshiki fabric, 1930s
   
   ADDITIONAL ATTRIBUTES: African cherubs with blingy rhinestone eyes, Kente cloth, 1990s Harlem, grainy vintage photograph

See results for the above text prompts below. All images were generated with Dall-E 3 (As of May/June, 2024). DOWNLOAD INITIAL COMPOSITION WORKSHEET

Evaluation is a critical aspect of assessing the quality and relevance of generated images. Following the initial composition of text prompts, students will engage in a critical assessment and revision process to address misalignment—that is, the degree to which generated images and their given text prompts are (dis)similar. This process involves examining the output to determine its coherence, accuracy, and alignment with the provided text prompt. This step is crucial, as a growing body of research on algorithmic bias demonstrates that misalignment often results in representational harms (harmful associations of specific traits with minoritized social identities) that disproportionately impact minoritized and vulnerable communities. For example, an AI-generated image might invoke visual elements that, while not independently racist, become harmful when applied to compositions depicting particular minoritized groups OR an AI-generated image might visualize elements that devalue/erase/white-wash artifacts, attributes or activities integral to the identities of a particular minoritized group.

The case studies below illustrate how text prompts, irrespective of whether they include demographic descriptions, might lead to the generation of images that reinforce unjust social hierarchies along intersecting axes of race, gender, disability, and geopolitics. Research has shown that such biases persist despite user interventions like carefully crafted text prompts or structural interventions such as algorithmic guardrails.

1. Case Study I: See image results by the artist, Minne Atairu’s whose text prompt for blonde braided hairstyles on richly melanated Black twins instead yields images of fraternal twins sporting wavy Blonde hairstyles.
   Generative Model: Midjourney, version 4
   
   
2. Case Study II: See image results by an Asian American MIT student, Rona Wang whose attempt to edit an existing portrait using the text prompt - “a professional LinkedIn profile photo” resulted in lighter complexion and blue eyes.
   Generative model: Playground AI
   
   
3. Case Study III: See images by researchers, Gautam, Venkit, & Ghosh (2024) whose text prompt ‘An upper class family’ generates a set of images depicting affluent, white families. As noted by the researchers, "While these images may conform to social norms in certain contexts, they inadequately represent the diversity of familial structures worldwide. The generated image underscores the inherent nature of stereotyping, wherein generalized beliefs about individuals’ personal attributes are formed based on their socio and demographic characteristics".
   Generative model: Imagen 2 by Google

Our refinement worksheet below will enable students to critically identify misalignments related to their text prompts, refine these prompts to address the identified concerns-if possible or even chose to abandon utilizing a generative system to achieve their artistic goal. Most importantly, the critique phases gives students the opportunity to understand the limitations of your selected text-to-image model. DOWNLOAD THE CRITIQUE WORKSHEET

AI DISCLOSURE
As AI-generated imagery becomes increasingly unrivaled in visual fidelity, and surpasses human perceptual capabilities⁸, we ask: how might guardians, administrators and all stakeholders identify AI-imagery generated by students? What steps might art educators take to ensure transparency and avoid misrepresentation to all stakeholders?

First, it is crucial students that students clearly disclose the synthetic nature of their works in either the artwork's title, its description, or both. See suggestions below:



AI WATERMARKS
Beyond captions and descriptions, which cater to human viewers, AI watermarks ensure that students' synthetic works are identifiable and traceable across the web, even after cross-platform modifications. Text-to-image models like Google's Imagen and OpenAI's DALL-E 3 automatically embed imperceptible watermarks directly into the pixel data of generated images.

To check the provenance of an image, use Maybe's AI Art Detector.

CLASS DISCUSSIONS & PRESENTATIONS
Avoid anthropomorphizing the technology, or obfuscating the people that utilize the tool. Instead, aim for sentences that clearly highlight the type of generative tool and the name of the student involved in making the work. This precise naming prevents misleading or vague references often seen with terms like "AI," "the AI," or "the Artificial Intelligence" that project human-like attributes (such as intent, agency and identity) to the algorithm, while shifting responsibility and accountability from developers of the technology. Below, we have created examples showing how we describe works. To reveal a sentence, click on an artist's name.