The four legitimate jobs; image prompt anatomy; iteration discipline; surgical edits.
Illustration · graphic design · animation · motion · 3D — 4 hours, drop-in friendly
Image Generation and Editing for Illustration, Design, and Previs
Use image tools the way professionals actually use them — as sketching, reference, compositing, and repair systems — with an iteration discipline that proves your authorship instead of eroding it.
01This Week at a Glance
You will make: a process board — 6+ iterations, keep/kill decisions with written reasons, one human-repaired direction.
Training-data tiers, output copyright, Content Credentials, disclosure.
Campaign poster: generate → reject → critique → repair.
Your process board, then defend-a-kill critique.
The reframe that makes this week work
The deliverable of an image session is not an image — it's a decision trail. Professionals generate wide, kill fast with reasons, and repair by hand. The rejects, and your reasons, are where your taste becomes visible and your authorship becomes defensible.
Opted out of generating?
Full critique/direction track today: you'll brief, judge, and document — the art-director round. Higher-order skills, same deliverable, equal credit.
02Key Concepts
1 · The four legitimate jobs (and the illegitimate one)
| Job | Looks like | Who uses it |
|---|---|---|
| Sketching/exploration | 20 thumbnails in 10 min to find what you don't want | Illustration, design |
| Reference & previs | Lighting studies, mood frames, set ideas | Motion, animation, 3D, video |
| Compositing parts | Generated texture under hand-made foreground | Design, illustration |
| Repair & extension | Inpaint a region; outpaint a new aspect ratio | Everyone in production |
The illegitimate job
Final art replacement. Not because a rule says so: it's the part that's actually yours, raw output isn't copyrightable, and clients can buy raw output without you. Generating your final art is competing with yourself at €0/hour.
2 · Image prompt anatomy (your Week 2 brief, visual edition)
- Subject + action — concrete, specific.
- Territory qualities — your verified Week 3 vocabulary: materials, light behaviour, colour logic, texture.
- Composition — framing, viewpoint, and a named empty zone with a purpose ("clear upper third for typography").
- Medium/process words — "gouache on rough paper" beats "painterly"; "35mm photo" beats "cinematic." Process words invoke physics; style words invoke the average.
- Avoid list — always includes "no text, no logos, no letters." Models grow fake type; type is yours anyway.
3 · Iteration discipline: wide, then narrow
- Generate wide — 4–8 from one brief-shaped prompt.
- Kill fast, with reasons — "Kill: texture is mush where the eye lands." One line each.
- Select — with a reason too.
- Refine narrow — change exactly ONE variable. Rerolling is a slot machine; controlled variation is direction.
4 · Surgical edits: the production skills
- Inpainting (Generative Fill) — regenerate only a selected region. Your hand chooses what changes.
- Outpainting (Generative Expand) — extend past the frame; reformat without re-shooting. Verify the invented regions.
- Image-to-image — your sketch in, rendered options out; strength decides how much of your structure survives. If the composition isn't yours anymore, you gave it too much freedom.
- Upscaling — invents plausible detail. Inspect at 100% before print.
5 · Rights, tools, and the classroom-safe choice
- Training tiers: Firefly trains on licensed/Adobe Stock content — the conservative default for classroom and client work. Most other tools train on scraped data; using them is a values decision to make knowingly. (Honest footnote: Firefly's contributor compensation is itself contested — "licensed" is cleaner, not spotless.)
- Output copyright: purely AI-generated images generally can't be copyrighted; your edits, composition, and combination are what's protectable.
- Content Credentials (C2PA): tamper-evident "how this was made" metadata. Provenance is becoming infrastructure.
- Disclosure: generated elements that survive into final work get disclosed, specifically.
03This Week's Words
Text-to-imageGenerating images from description. A sketching tool, not a finishing tool.
Image-to-imageYour image + instructions → variations. Strength = how much of yours survives.
InpaintingRegenerating a selected region only. Surgical, not slot machine.
OutpaintingExtending beyond the original edges. Reformatting's best friend.
UpscalingResolution increase with invented detail. Inspect before print.
SeedThe random start number. Same seed + prompt ≈ same image — controlled variation.
Structure controlSteering with a sketch/pose/depth map: composition obeys, rendering varies.
Negative promptThe avoid list. "No text" is almost always in it.
Latent spaceThe model's map of visual concepts — prompting is navigation, not retrieval.
Content CredentialsEmbedded provenance metadata (C2PA) — machine-readable "how this was made."
04Demo: Campaign Poster — Generate, Reject, Repair
I made this, so can you
A poster background for a contemporary dance festival ("Hinge") taken through the full professional loop.
Prepared demo artifact
Instructor backup / finished example: process-board.html. Use it if the live tool run fails, or to reveal the finished structure after critique.
- Brief-shaped prompt: territory qualities + composition with a named type zone + palette + "no people, no text."
- Rejection round: 4–6 outputs, keep/kill + one reason each, fast and aloud — and the room overrules one call.
- Critique loop: the survivor screenshot into a chat model with the brief — "where will typography fight the image? what reads as AI-generic at arm's length?"
- Human repair: crop for the type zone, generative-fill ONE region, levels, real type placed. Five minutes of recognizable design craft.
The planned failure
At least one generation ignores "no text" and grows fake letters. It always does. That's why type stays yours.
05Step-by-Step Workflows
From prompt to process board
- Write the image brief from your Week 3 territory (or build a quick territory first).
- Generate 6–8. Number them. No judging mid-generation.
- Kill round: minimum 4 killed, one written reason each — specific beats vibes.
- Select + reason. One survivor (two max).
- One controlled refinement: change exactly one variable; note what and why.
- One surgical edit — inpaint/crop/adjust, by hand where you can.
- Board it: iterations in order, decisions visible, final direction + repair notes. (Process documentation template)
Image-to-image from your sketch
- Rough sketch (90 seconds, ugly is fine) of your composition.
- Photograph/scan → image-to-image with your territory prompt.
- Strength high enough that your composition survives.
- Treat results as rendering studies of your drawing — steal back what works into the real piece.
06Prompt Recipes
The image brief[Subject + action, concrete]. Visual territory: [3–5 verified qualities — materials, light behaviour, colour logic, texture, process era]. Composition: [framing/viewpoint] with [empty zone + what it's for]. Medium: [process words — e.g., flat screen-print shapes / gouache on rough paper / 35mm photograph]. Palette: [named colours]. Avoid: text, logos, watermarks, [your clichés].
The variation directorKeep everything about option [N] the same, but change only: [one variable — the light direction / the camera height / the palette to X]. Do not change the composition.
The multimodal critiqueCritique this image against my brief: [paste brief]. Where will typography fight the image? What reads as AI-generic at arm's length? What would a print production person flag? Don't soften.
The repair plannerI'm keeping this image but repairing it by hand. Based on my brief, list the 5 highest-impact manual interventions (crop, paint-over, colour, type zone, texture) in priority order, with one sentence on why each earns its place.
07Lab: From Prompt to Process (70 min)
Cold start friendly — a quick territory takes 10 minutes if you missed Week 3.
Your subject, by track
IllustrationBook cover rough for a folk tale you actually like — territory from its setting.
Graphic designGig/event poster background with a named type zone.
Animation / motionThree style frames for one shot — same scene, three light moods.
3DPrevis of a set/environment you might actually build — composition + lighting study.
The structure (all tracks)
- Image brief from your territory — 10 min
- Generate 6–8 — 10 min
- Kill round with written reasons — 10 min (the core assessment artifact)
- Select + one controlled refinement — 10 min
- One surgical edit or crop by hand — 15 min
- Board assembly + selection notes — 15 min
Zero-generation track
Art-director round: you write the brief and make every keep/kill call; a partner operates. Or run the full kill round + critique + repair plan on the instructor's prepared 8-image set. The decision trail is the deliverable either way.
Free generations ran out?
By design: documentation, critique loop, and repair planning on what you already have completes the deliverable. Nobody is blocked by a quota.
08Integrity Check
This week's four
- Could you show this board in a crit and defend every kill?
- Is anything on the board secretly "final art" you're hoping to ship raw?
- Did you choose your generator knowingly (training data, rights tier) — or by default?
- Write the disclosure line now, while the process is fresh.
09Troubleshooting
| Problem | Fix |
|---|---|
| Everything looks "AI" (waxy, over-lit, hollow) | Add process/medium words and texture qualities; ban the defaults in your avoid list |
| It keeps adding text/letters | It will. "No text, no typography, no letters" — and plan type as your job |
| Same-y results across prompts | You're rerolling, not directing. Change one named variable — or your territory is too thin; go back to qualities |
| Hands/faces/details broken | Inpaint the region, crop past it, or let it not matter — it's a rough, not a final |
| Free generations ran out | Switch to kill-round documentation, critique, repair planning — that's most of the deliverable |
| "It's doing the fun part" | Keep the fun part: sketch first, image-to-image at high structure strength — outputs become rendering studies of YOUR drawing |
10Resources
- VideoHow AI Image Generators Work — ComputerphileEmbedded above in Concepts. The diffusion mechanism, explained at a whiteboard — the magic and the mush both make sense afterwards.
- DocsLinkedGenerative Fill — Adobe Photoshop documentationThe official reference for the inpainting workflow in today's demo and lab.
- DocsVerify before teachingMidjourney documentationParameters, image prompts, style references — useful even if you only watch demos. Features shift fast.
- StandardLinkedContent CredentialsThe C2PA provenance standard — how "what made this image" becomes verifiable metadata.
- OfficialLinkedCopyright and AI — US Copyright OfficeThis week's angle: registration guidance for works containing AI material — your human contributions are what's protectable. Document them.
- ToolLinkedHave I Been Trained — SpawningRe-check from Week 1, now that you're generating: see the training sets, register opt-outs for your own work.
- VideoVerify before teachingAdobe Firefly official tutorials (Adobe Learn / YouTube)Short official walkthroughs of Fill/Expand — the UI changes often; grab the current one the week you teach.
Optional extension
- Take one rejected generation; write 3 sentences on what would have to change to pass your rubric.
- Run the same brief in Firefly and one other tool; note 3 differences in their defaults — different training data, thinking differently.
- Hard mode: recreate your selected direction entirely by hand (sketch/collage/photo). What did each version teach you?
Next week: Motion Design → — your image skills become style frames; AI starts planning shots and production tasks.