A prompt is a draft. Generative imagery is a conversation, not a search.
By the end of Day 1, students have a working literacy for how generative image systems behave — not the math, but the moves. They see how the same prompt can become different images, how one word changes the result, how model defaults and biases show up, and why iteration is the honest way to use these tools. Along the way, each student makes two documented image studies as evidence of what they noticed.
Schedule
Block 1: instructor demo, class co-construction, student studio time
Block 2: second lesson sequence or guided production sprint
Break: 15 minutes away from the screen
Block 3: studio work with instructor/tutor nudges
Block 4: finish, export, document
Daily share-back: every student shows one decision
Reflection card and next-day preview
Lessons
Lesson 1.1 - First prompts
Generate, judge, and iterate. Notice what changes when you change one word.
I do / We do / You do
I do: 10 min: generate one image, make three flawed iterations, and narrate what changed.
We do: 15 min: class co-writes a prompt and one student drives the revision.
You do: 25 min: each student makes 5 images on a self-chosen theme and names a favorite.
Deliverable: 5-image grid plus one paragraph, "what I noticed."
Stretch: Same theme, three different style constraints inside the available image tools.
AI tutor aiGuidance excerpt
- nudge: If the student keeps the first image, ask what they would change.
- nudge: When a student uses a vague adjective, ask for two more specific words.
- require: Do not generate or accept prompts that describe real, identifiable people.
Lesson 1.2 - Style and specificity
Learn to constrain. Use references, not vibes.
I do / We do / You do
I do: 10 min: compare painter, photographer, and design-discipline references.
We do: 10 min: class chooses one constraint and compares results across seats.
You do: 25 min: each student regenerates one image three more times with added specificity layers.
Deliverable: 4-image series, original plus 3 specificity layers, with a one-sentence style statement.
Stretch: Write a negative prompt and document what it changes.
AI tutor aiGuidance excerpt
- nudge: Ask what visual evidence would prove the style constraint worked.
- nudge: Push students from mood words toward observable details: lens, surface, palette, composition.
- require: Cite any artist or named work used as a reference.
How the tech works
An image generator is not searching the web for a picture. It starts from a learned relationship between language and images, then produces a new output that statistically fits the prompt. For students, the useful model is functional: the text prompt steers a denoising process, the image is a proposed answer, and the next prompt is a revision. Stanford CRAFT and MIT Sloan both frame AI literacy around concrete probes and iterative prompting rather than memorized formulas (Stanford CRAFT; MIT Sloan prompt guidance). The control points students can feel today are specificity, constraints, references, and negative space: what to include, what to avoid, and what visual evidence would count as success. The iteration loop runs between the student, the prompt, and the generated image. A weak first image is not failure; it is diagnostic material. That is why the instructor demos bad first outputs on purpose. Students also learn that the model has no classroom judgment: it may satisfy the grammar of a prompt while missing the intent. The human job is to diagnose what is missing, add constraints, remove ambiguity, and decide when an output has become specific enough to keep. This is why the first deliverable is a grid with notes, not a single polished image. The process record is part of the work.
A short history
The modern image-generator moment sits on several earlier AI landmarks. AlexNet's 2012 ImageNet win made deep-learning vision impossible to ignore; GANs in 2014 showed that neural networks could learn to generate plausible images; diffusion research in 2015 helped establish the denoising family of approaches that later image systems popularized. By 2022, DALL-E 2, Stable Diffusion, and Midjourney made generated imagery visible to non-specialists at public scale, which is why Day 1 treats image generation as both a studio tool and a media-literacy problem. This short history also lets students see that today's interface is not isolated from older art questions. Artists have always used tools, chance operations, references, and systems; what changed is the speed and scale of synthesis. The useful question is not whether the tool is new, but what responsibilities arrive when a student can produce plausible images quickly.
Ethics for this day
Day 1 ethics starts with bias and consent. The class runs a simple Stanford CRAFT-style probe: ask for "a CEO," "a nurse," and "a janitor," then compare demographics and visual defaults (Stanford CRAFT). The deeper risk is not awkward stereotypes but targeted harm. The 2022 Westfield, New Jersey high-school deepfake incident and Thorn's 2024 research on teen awareness of "deepfake nudes" make the code of conduct concrete: no imagery of real classmates, teachers, or identifiable minors; no sexualized imagery; and no content that would violate school policy or law. The policy is stated before production because students need the boundary while the tool is in their hands, not after something goes wrong (Thorn research; Common Sense Media).
Real-world applications
Editorial illustration roughs
Concept art and moodboards
Set and prop design references
Fashion and product mockups
Campaign image directions
Famous works to study
Refik Anadol, Machine Hallucinations (2017-). Large-scale data paintings that translate image archives into immersive generated visual fields. Source.
Memo Akten, Learning to See (2017-18). A live camera system that reinterprets the world through learned visual categories. Source.
Refik Anadol, Echoes of the Earth: Living Archive (2024). Serpentine installation using environmental data and AI-generated imagery. Source.
Boris Eldagsen, Pseudomnesia: The Electrician (2023). AI-generated image that won, then was declined from, a Sony photography award; included here as a study in authorship controversy. Source.
Holly Herndon, Holly+ (2021-). Voice-avatar project exploring consent and attribution around AI voice models. Source.
What you should be able to do by the end of today
Generate a set of related images from one theme.
Name which prompt changes affected the output.
Use references and constraints without copying blindly.
State the Day 1 risk policy in your own words.
Deeper look
Optional note for the curious
Diffusion is different from a chatbot because it does not write one word after another. In the classroom version, think of it as repeatedly cleaning up visual noise until the image fits the text condition. The exact architecture matters less than the studio behavior: small changes in wording can steer the denoising path toward a different composition.