Where we start
This course runs on one job. Daybreak, a coffee subscription company, is launching its autumn roast and needs six illustrations that look like one hand made them: an email hero, a landing-page header, three social tiles and the sleeve artwork. The sleeve carries an origin claim that legal signs, which is the first reason no lettering may come out of a model for this job. Every session moves those six assets one step closer to shipping.
★ What you walk out with tonight. A one-page control sheet for one of the six assets: the model and version with a date, the ratio, the reference handles, the seed and what it is worth, and a line naming what will be set by hand afterwards. Plus the vendor sentence that tells you why each control is only as strong as it is.
Four neighbours, so you land in the right place. Copyright, indemnities, the court cases and provenance as a signature are learn-ai-design-with-phoebe a4. Disclosure duties and the EU dates are learn-ai-media-with-phoebe session 6. Constraint briefs for product UI are learn-ai-design b2 and b7. On-brand image recipes for marketing volume are learn-ai-marketing b8. Here the subject is illustration craft: the picture before the style, the type kept out, consistency as a system, and the judgement to say no.
Five layers, one of them a sentence
5 min liveRead the parameter list of any current image model and the controls sort themselves into layers. Some fix a property of the output outright. Some nudge it. One layer is words, and words nudge. Knowing which layer a problem lives in is the difference between fixing it and typing "more cinematic" for the fourth time.
LiveBinds, persuades, and the test that tells them apart3 min▶
A control binds when the output cannot violate it. Ask for 3:2 and you get 3:2; the vendor's page even warns that "aspect ratio isn't the same as image dimensions", which is the sound of a control that is exact about one thing and silent about another. A control persuades when the output usually honours it and sometimes does not, with no error raised. Every word of a prompt is in that second category, because the text encoder turns the sentence into a weighting and the picture is a negotiation between all the weights at once.
The test: run it ten times. A binding control gives the same property ten times. A persuading control gives it seven or eight times and you will not know which two until you look. Layers 1, 3 and 5 pass the test. Layer 2 passes in proportion to its weight. Layer 4 does not, and it never will, because that is what an encoder is.
This is why "make it look like our brand every time" cannot be answered in layer 4. It is a question about layers 1, 2 and 5, and session 5 answers it there.
Self-studyWhy the model version is layer 1, and how fast it moves4 min▶
The version decides which controls exist. On Midjourney the default became V8.2 on 24 July 2026, and that release folded Omni Reference, Character Reference and the Retexture tool into one Edit Model that takes up to four reference images. V8.1 had been the default only since 10 June; V7 before that since June 2025. Three defaults in fifteen months, and each one retired or renamed a control that prompt guides were still teaching. A brief written for V7 with --oref in it is talking to a parameter that the current parameter list marks "replaced by the Edit Model in V8.X".
The same is true across vendors. OpenAI's image models are named gpt-image-2.5-sunburst and -flare and the guide says which is for editing precision and which for speed. Google's are gemini-3.1-flash-image and gemini-3-pro-image, and the number of reference images each accepts is different (up to fourteen on the lightest, six object and five character and three style on the Pro). None of these names will be current for long. The professional habit is to put the model and version in a dated footnote on every control sheet and never in a headline, because the headline will be wrong within a quarter and the footnote will tell you why.
Daybreak's spring campaign was built on a saved prompt with a character reference flag in it. When the summer default changed, the flag stopped doing what it had done and nobody noticed for three assets, because a persuading layer does not throw an error. It just draws a slightly different mug. The fix was one line at the top of the brief: the model, the version and the date it was checked.
What each control is documented to do, in the vendor's words
6 min liveThe most useful sentences in any vendor's documentation are the ones that say what a control cannot do. They are always there and never in the marketing. Six of them, one per control, read on 18 September 2026.
| Control | Documented range | The vendor's own limit | Layer |
|---|---|---|---|
| Aspect ratio --ar | up to 14:1 on V8.x (4:1 for HD) | "Aspect ratio isn't the same as image dimensions" | 3, binds |
| Seed --seed | any integer | "(99% identical)" on V8.1 and V8.2 | 3, binds at 99% |
| Style reference --sref + --sw | weight 0 to 1000, default 100, V6 and later | Style only, "not objects or people"; no code from an uploaded image | 2, binds by weight |
| Edit Model (V8.1+) | up to 4 reference images | Not compatible with --tile; combine characters into one image when features mix | 2, binds by weight |
| Omni reference --oref (V7 only) | --ow 1 to 1000, default 100 | One image; 2x GPU time; keep below 400 unless stylize is very high | 2, retired in V8 |
| Output size (OpenAI) | 1024 square, 1536x1024, 1024x1536, or multiples of 16 between 1:3 and 3:1 | "can still struggle with precise text placement and clarity" | 3, binds |
| The prompt | any words | every vendor: none of the above is promised by words alone | 4, persuades |
Read 2026-09-18 from docs.midjourney.com (Version, Parameter List, Style Reference, Edit Model, Omni Reference, Aspect Ratio), developers.openai.com and docs.ideogram.ai. Every one of these moves; the date is part of the fact.
LiveReading a parameter page like a contract3 min▶
Three habits turn a docs page into a control sheet. Find the range. A control with a stated range (0 to 1000, up to 4 images, 14:1) binds within it; a control described only in adjectives persuades. Find the parenthesis. The most important fact on Midjourney's version chart is two words in brackets after the word Seed, and they are the reason session 5 exists. Find the "cannot". Every page has one: cannot make a code from an upload, not compatible with tile, may struggle with text placement. The cannot is the boundary of the control, and it is always more specific than the can.
The fourth habit is the one people skip: write down the date you read it. Three defaults in fifteen months means a control sheet without a date is a control sheet about a model that may no longer exist.
The seeded prompt has one binding control, the ratio at the end, and fourteen words of persuasion. Seven of the fourteen are about quality and none of them names an object, a light or a colour. Session 4's bench will count exactly that.
Daybreak's launch set, as a control problem
4 min liveSix pictures that must look like one hand made them, delivered across an email, a web page, three square tiles and a printed sleeve. Written as a control problem, the six assets share three layers and differ in two, and the sharing is what makes them a set.
LiveWhat a set is, in one sentence2 min▶
A set is six pictures that share their binding layers. Same model and version, same reference kit at the same weight, same palette named in words, same rule that type is set afterwards. What differs is the ratio and the destination, and both of those are binding controls too, so nothing about the set depends on a persuading layer holding. The prompt for each asset is then short, because the picture has already been decided three layers up.
The sleeve is the asset that makes the rule non-negotiable. It carries an origin claim that legal has signed off word by word. A model that spells it wrong once in a hundred has spelled it wrong once, on a printed bag, with a lawyer's name on the approval. That is why layer 5 is where the words live, and session 4 is the whole mechanism behind it.
Self-studyWhy the leaders' question and the practitioners' question are the same question3 min▶
A head of marketing asks "can we make it look like us every time?" A designer asks "which flag do I use for consistency?" They are asking about the same layer. The honest answer to both is that no vendor sells identity-level repeatability as a feature; two of them document the gap in their own guides, and the third puts "(99% identical)" in brackets next to the control everyone assumes is exact. The answer is a pipeline, and the pipeline is a leadership decision about what gets locked and a practitioner's discipline about what gets recorded. Session 5 is that pipeline. Tonight's control sheet is its first page.
When Daybreak's designer was asked why tile 3 looked "a bit off", she opened the control sheet and read the date: the sheet said V8.1, the tile had been regenerated in the week V8.2 became the default. Same prompt, same seed, same references, different encoder. The sheet did not prevent the drift. It made the cause a fact instead of an argument.
A control sheet for one asset
★ 22 min · everyone writesPick one of the six, or one asset from your own work. One page, five layers, one line each, and the rule for the whole exercise: every line names a control you can point at in a docs page, with the date you read it. The prompt itself is the last line and the shortest.
Layer 1: name and date. The model, the version, the date you checked the version page. If you cannot find a version page, write that down too; it is a fact about the tool.
Layer 2: references. Which images the model copies from, what kind of reference each is (style, subject, composition), and the weight. If you have no references yet, write "none, so consistency is unmanaged" and underline it.
Layer 3: the binding numbers. Ratio. Seed, with "99%" written next to it. Any flag with a stated range, and the range. Nothing described in adjectives goes on this line.
Layer 5: what is hand-set. Every word that will appear in the finished asset, and the sentence "set in HTML or SVG after generation, not requested from the model." For the sleeve, add who signs the words.
Layer 4: the prompt, last. Subject, setting, light, palette, one style. Under forty words. If it is longer, something from layers 1 to 3 has leaked into it as an adjective.
Head the page with the asset's job, not its name: "email hero, first thing a subscriber sees, 3:2, headline set over the top third." The job decides the ratio and the zone before anything is generated.
Open the vendor's version page and copy the default version and its date into layer 1. Do not write it from memory; the point of the exercise is that memory is what drifts.
List the references you actually have. A moodboard is not a reference kit. A reference kit is named files with a purpose each, and session 2 builds one.
Write the ratio and the seed, and next to the seed write what the docs say it is worth. On the current default that is two words in brackets.
Draw the type zone as a dashed rectangle on a thumbnail of the asset, and write the words that will go in it. Those words never enter the prompt.
Now write the prompt, and count its words. Bring the count, and the number of words in it that name a colour, a light or an object, to session 4.
If you take one thing from tonight. The prompt is not the work. The prompt is the last and weakest of five controls, and art direction is the other four. A control sheet with a dated version, a reference kit, a recorded seed and a type zone has done more than a hundred prompt revisions.
Before session 2
25 min- Finish the control sheet for one real asset from your own work, all five layers, with dates. Keep it; session 5 adds a column to it.
- Read one vendor's parameter page end to end and write down every "cannot", "not compatible" and parenthesis you find. Most people find five or more they had never seen.
- Collect five images that look like your brand, from anywhere, and for each write one sentence on what makes it look that way: a colour, a line weight, a light, a texture. Session 2 turns those sentences into style DNA.
- Sort your last three prompts into layers. Count the layer 4 words. Write one sentence on what you were trying to make a persuading layer do that a binding layer should have done.
Three questions
Clear all three and session 1 gets stamped in your passport.
1 · A designer regenerates an approved asset with the same prompt and the same seed and it comes back slightly different. What happened, most likely?
The compatibility chart says it in two words. A seed on V8.1 and V8.2 reproduces at ninety-nine percent, and if the default moved between runs the encoder itself changed. Neither is fixed by more words in layer 4. The control sheet's date is how you tell which one it was.
2 · Which of these is a binding control, and which persuades?
The test is ten runs. A ratio gives the same ratio ten times. A reference at weight 100 gives its style in proportion. The word "amber" gives amber most of the time, and the two times it does not, nothing tells you. That is what a text encoder is, and it is why the type is set by hand.
3 · Why does a control sheet carry a date on the model version?
V7 to V8.1 to V8.2 inside fifteen months, with Omni Reference and Character Reference folded into an Edit Model along the way. A brief that says "--oref" is talking to a retired control and the model will not tell you. The sheet's date is the only way to know which model made which asset.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map.
Session 1 cheat sheet · pin this
Your control sheet has a layer 2 line that probably says "none yet". Session 2 fills it: style written down in words that outlive the person who wrote them, and the reference kit that turns "looks like us" from an opinion into files with names. Bring the five images and the five sentences.