Where we start
Session 1 ended with a control sheet whose layer 2 line says "none yet", and a folder of five images that look like your brand. They look like it for reasons nobody has written down, which is the exact state Daybreak was in when its spring designer went on leave and three autumn assets came back "sort of ours". The six launch pieces still have to look like one hand made them: the email hero at 3:2, the page header at 21:9, three square tiles and the printed sleeve. The approved look is a ceramic mug of black coffee on a worn oak counter, morning light from a window on the left, warm amber and burnt orange and cream, flat vector with the grain of the paper showing. Tonight that look gets written down, and the "none yet" line gets filled with things a stranger could act on.
★ What you walk out with tonight. A six-field style sheet for your own brand: line, colour, light, texture, composition habit and what is absent, one sentence each. And a named reference kit: four files, one purpose each, a style weight and a date. Together they fill layer 2 of the control sheet, and they keep working when the person who wrote them is not in the building.
Three neighbours, so you land in the right place. Constraint briefs for product UI and --sref as reference discipline for screens are learn-ai-design-with-phoebe b2 and b7. On-brand image recipes at marketing volume are learn-ai-marketing b8. Here the subject is illustration style written down so that it outlives the person who wrote it: the sentences first, the files second, the weight recorded.
Six sentences a stranger could act on
6 min liveA style sheet is a form with six fields, and each field takes one sentence. The test for every sentence is the same: read it to somebody who has never seen your work, show them three images, and ask them to point at the one that matches. If they cannot point, it is not a sentence yet. Here is Daybreak's, filled in.
LiveWhy a style has to be sentences before it is references3 min▶
An image reference carries everything about the image. The mug, yes, but also the angle you did not choose, the shadow the photographer did not notice, the slightly warm white balance, the handle pointing left. The model copies from all of it, in proportion to a weight, with no idea which parts you meant. A sentence carries only the choice. "One window, from the left, morning" says nothing about the mug and everything about the light, and a stranger can check it against any output in two seconds.
The six fields are the six choices a generator can actually be pointed at: line, colour, light, texture, composition habit, and what is absent. Each takes one sentence. If a field needs the word "nice" or "premium" to get written, it is not a choice yet; it is a mood, and moods belong in layer 4 where they persuade and sometimes lose.
The sixth field does the most work, because it is the one no reference image can express. A picture cannot show the absence of people. It can only show one picture that happens to have none, and the model has no way of knowing that the absence was the point. "No people, no text, no logos" is a sentence a stranger can check on every asset, and it is the kind of sentence that would have saved Daybreak's spring look.
Self-studyWhat the four anchor kinds are, and why the bench counts them4 min▶
Session 4 runs a bench over a brief. No image is generated; the text is counted, because the text is the one input a person fully controls. Among the things it counts are concrete anchors, drawn from four lists in its source: colour words, light words, placement words and material words, plus numbers and capitalised proper nouns. The lists are deliberately finite. A colour not on the list is not counted, and the learner can read the list.
Map them onto the sheet. Colour words are the Colour field. Light words are the Light field. Placement words are the Composition habit field. Material words are Texture and, mostly, Line. The style sheet is where those words are chosen once, so every brief written from it inherits them, and a brief written from a good sheet starts with a high anchor count before anyone has described the subject. The verified ladder shows the gap: the first prompt scored zero anchors and seven style words; the concept brief scored fifteen anchors and two.
What the bench cannot do is judge the sixth field. It counts a no-text clause as lettering refused, and that is the only absence it looks for. The rest of what is absent lives in the sheet because a person has to hold it.
Daybreak's spring look lived in one designer's head. When she went on leave, three assets came back "sort of ours" and nobody in the room could say what was wrong. They shipped two of them. When she returned she looked for four seconds and said: "the light is from the right, and there is a gradient." Those two sentences were the whole style, and they had never been written down.
Three kinds of reference, three things they carry
5 min liveOnce the sentences exist, references stop being "pictures I like" and become files with a payload. There are three payloads: how it looks, what is in it, and where things sit. Every vendor names them differently and caps them differently, and the caps are the useful part.
| Reference kind | Carries | Vendor control | Stated limit |
|---|---|---|---|
| Style (Midjourney) | look and feel | --sref + --sw 0 to 1000, default 100, V6 and later | Style only, "not objects or people"; no code from an uploaded image |
| Subject (Midjourney) | the mug, the counter | Edit Model, V8.1 and V8.2 | Up to 4 reference images; replaces Omni and Character Reference; fine details "may not perfectly match" |
| Composition (Adobe Firefly) | "outline and depth" | Composition reference, Strength slider | Strength sets how far the outline is followed; a slider, not a lock |
| Style (Adobe Firefly) | "look and feel" | Style reference, Strength slider | Named for brand consistency; the same slider, the same reservation |
| Object, character, style (Google Gemini) | all three, capped | Per-model reference ceilings | Up to 14 total on the lightest model; 6 object, 5 character, 3 style on the Pro |
Read 2026-09-18 from docs.midjourney.com, helpx.adobe.com and ai.google.dev. Every one of these moves; the date is part of the fact.
LiveStyle, subject, composition, and which vendor names which3 min▶
Midjourney's style reference is the cleanest statement of the first kind. --sref points at an image or a code, --sw runs from 0 to 1000 with 100 as the default, it works on V6 and later, and the page says what it carries: style, "not objects or people". Subject continuity is a different control. On V8.1 and V8.2 the Edit Model takes up to four reference images and, in the version page's words, replaces Omni Reference, Character Reference and the Retexture tool. Two controls, two payloads, and mixing them is the commonest reason a "style reference" of the mug produces a different mug.
Adobe Firefly names the third kind. Composition reference matches "outline and depth" and style reference guides "look and feel" for brand consistency, each with a Strength slider (pages updated 2026-06-16). Google caps references per model rather than per kind: up to fourteen total on the lightest Gemini image model, and on the Pro six object, five character and three style. Same three payloads, four vocabularies. The style sheet is written in none of them, which is the point: it survives a change of vendor, and the vendors change.
Self-studyWhat no reference can do3 min▶
Read the cannots. A style code cannot be made from an uploaded image, in Midjourney's own line, so a code and an image are different handles even when they produce the same look. --sw is not compatible with Moodboards. Omni Reference on V7 was one image at twice the GPU time, and the parameter list now marks it replaced by the Edit Model in V8.X. On subject continuity the vendor's own wording is that fine details "may not perfectly match".
Put those together and the shape is clear. A reference binds by weight: at --sw 100 the style arrives in proportion, at 1000 it dominates, at 0 it is off. Nothing on that dial is identity. Four images of the mug make a mug that is close and, the vendor says, may differ in detail. That is why session 5 treats consistency as a pipeline with a human comparing outputs, and why tonight's kit records the weight instead of trusting the default to stay put.
The reference kit: named files with a job each
4 min liveFive images came in the door. Four go into the kit, and the fifth goes back in the folder. What makes the four a kit is not that they are the best pictures; it is that each one has a name, a payload, and a reason to be there that can be read off the style sheet.
LiveFour files, each with one job3 min▶
A kit is not a moodboard. A moodboard is everything you like, and it is useful for exactly one afternoon: the one where the six sentences get written. After that it is noise, because a model pointed at forty images copies from forty images, and the weight is spread across every choice you did not make. A kit is the smallest set of files that, at a stated weight, reproduces the six sentences. For Daybreak that is four.
Each file is named and each has one purpose. DB-01 carries the approved mug and counter, and it is the subject reference: the Edit Model's input, not --sref's. DB-02 carries the light and the palette, DB-03 the paper grain close-up, and both are style references. DB-04 is a composition with the empty upper third, the one the type will own in session 3. There is at least one file of each kind, and no file is asked to carry two.
The weight is written down because the default is 100 and someone will change it. On the day they do, "the kit at 100" and "the kit at 300" are two different looks with the same four file names, and only the sheet can tell you which one the approved asset came from.
Self-studyHow the kit fills layer 2 of the control sheet3 min▶
Session 1's control sheet had a layer 2 line that said "none, so consistency is unmanaged". It now gets four file names, one weight, and one sentence: "style only; subject continuity via the Edit Model with DB-01." That sentence separates the two payloads on paper, so that nobody later feeds the mug to --sref and wonders why the mug moved.
When the version changes, and the version page says it has three times in fifteen months, the kit is re-run once at the stated weight and the outputs are compared to the six sentences, not to memory. "Light from the left" is either true of the new output or it is not. "Looks a bit different" is not a finding; "the texture field broke, there is a gradient" is. The sentences turn a re-check into a checklist, which is the whole reason they were written before the files.
Session 5 adds a column to this line: the seed, and what the seed is worth. Tonight's job is only to make sure layer 2 no longer says "none".
Six lines is the target, not a minimum. If a field needs two sentences, one of them is a second field wearing a costume, or it is a mood. The stranger test in the build-along tells you which.
The style sheet and the kit, for your brand
★ 22 min · everyone writesLay out the five images and the five sentences you brought. In twenty-two minutes they become a six-field sheet and a four-file kit, and the rule for the whole exercise: every sentence gets read aloud to a neighbour, and a sentence they cannot point at is rewritten before it goes on the sheet. The weight is the last thing written.
Six sentences. Write the six fields from the five images: line, colour, light, texture, composition habit, what is absent. One sentence each, under fifteen words, naming something a stranger could check. Start with the sixth; it is usually the easiest to see and the hardest to say.
Test each on a stranger. Read one sentence to your neighbour and ask them to point at the image that matches. A sentence nobody can point with is rewritten on the spot. "Warm" fails this test every time; "amber, burnt orange, cream" passes it.
Pick four files. Choose four of the five images as the kit and drop the fifth. The one to drop is the one that carries the fewest of your six sentences, not the one you like least. Name the four with a prefix and a number.
Give each one job. Next to each file name write its purpose: subject, style or composition. There must be at least one of each. A file that wants two jobs is two files, or it is the one you drop.
Write the weight. On the control sheet, next to the kit, write the style weight (100 unless you have a reason and can say it) and today's date. Then copy the line: "style only; subject continuity via the Edit Model with the subject file."
Cross out the moods first. Any sentence you brought with "nice", "premium", "clean" or "vibe" in it goes; those are layer 4 words and they cannot be pointed at. What is left is the raw material for the six fields.
Fill the six fields in order: line, colour, light, texture, composition habit, what is absent. If you cannot fill one from the five images, write "not decided" rather than inventing it. A gap on the sheet is a fact; a guess on the sheet is a future argument.
Run the stranger test on every line. Your neighbour points or they do not. Rewrite until they point at the right image, then move on. Do not explain the sentence to them; the explanation is what the sheet is supposed to replace.
Choose the four files by counting how many of the six sentences each image carries. Keep the four with the highest counts. Write the prefix and number on each: your brand's initials, then 01 to 04.
Assign one payload per file and check the set: at least one subject, one style, one composition. If your five images were all the same kind, write down which kind is missing; that is what the homework goes looking for.
Fill layer 2 of the control sheet: four file names, the weight, the date, and the one sentence that keeps subject and style apart. Read the line back. If a stranger could run it, you are done.
If you take one thing from tonight. The sixth field, what is absent, is the sentence no reference can carry and the one that most often names the whole style. Daybreak's spring look was two absences: no light from the right, no gradient. Write yours before you choose a single file, because it is the one thing the files cannot tell you.
Before session 3
25 min- Finish the six sentences and the kit for one real brand, yours or a client's. Every sentence passed by a stranger, four files named, one weight, one date, all on the control sheet's layer 2 line.
- Run the kit once at the stated weight if you have a tool, with the shortest prompt you can write. Then write one line on which of the six sentences the output broke. If none broke, write that down too; it is rarer than it sounds.
- Find one asset in your organisation that "looks off" and name the field it breaks. Not "the colours are wrong" but "Colour: there is pure black in the shadow." Bring the asset and the sentence.
- Bring one composition you admire, from anywhere, illustration or photograph. Session 3 is about where the picture sits and where the words will go, and it starts with your example.
Three questions
Clear all three and session 2 gets stamped in your passport.
1 · Why is the style written as six sentences before a single reference image is chosen?
A model pointed at an image copies the angle, the white balance and the handle direction along with the light you meant, in proportion to a weight it does not understand. "One window, from the left, morning" carries one choice and a stranger can check it in two seconds. And no picture can show the absence of people; only a sentence can.
2 · A style reference is attached at --sw 100. What does it do, and what does it not do?
The vendor's page says style only and not objects or people, gives the range and the default, and adds that a style code cannot be created from an uploaded image. Subject continuity is a different control with its own cap of four images and its own reservation about fine details. A weight is a dial, and identity is not on it.
3 · What separates a reference kit from a moodboard?
Forty images at one weight spread the model's attention across forty sets of choices you did not make. Four files with a name and a payload each, at a recorded weight, can be re-run after a version change and checked line by line against the six sentences. The date is what makes that check a fact rather than a memory.
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 2 cheat sheet · pin this
The six sentences say how the picture looks. Session 3 is where it sits: subject, light and ratio treated as layout, and the empty third that the type will own before any type exists. Bring the composition you admire and the DB-04 of your own kit, because that file is about to become a plan.