Where we start
The style sheet from session 2 has six fields: line, colour, light, texture, composition habit, and what is absent. The reference kit beside it has four named files, DB-01 to DB-04, at style weight 100. Together they say how Daybreak's autumn roast pictures will look. Nothing yet says where anything sits, and where things sit is what a ratio, a destination and a headline are about. Composition is decided before generation because it is the layer that decides whether there is room for the words. Six assets, six ratios, six empty zones, and one approved picture underneath all of them: a ceramic mug of black coffee on a worn oak counter, steam rising, three roasted beans and one green Guji bean beside it, morning light from a window on the left.
★ What you walk out with tonight. Six thumbnails, one per asset, each with its subject placed and its type zone dashed, and six placement lines written only in words the bench counts as anchors: where the mug sits, where the light comes from, what it stands on, and which third stays quiet. The composition you brought becomes the first of them.
Two neighbours, so you land in the right place. Why models cannot spell, and the bench that reads a brief for it, are session 4 and are not restated tonight. Slide and deck layout is learn-visual-storytelling-with-phoebe. Here the subject is composition as an instruction to a model, and the empty third as the most important instruction in it.
Six frames, six empty zones
5 min liveThe same mug goes to an email, a web page, three square tiles and a printed sleeve. Each destination fixes a ratio, and each ratio leaves a different part of the frame free for words. Drawn side by side at their true proportions, the six assets stop being six pictures and become one picture placed six ways.
LiveThe ratio is the first composition decision3 min▶
The ratio binds, which session 1 established, and it is set per destination, not per taste. An email hero is 3:2 because that is the width of a mail client with room for a headline over it. A landing header is 21:9 because a browser is wide and short. A social tile is square because the platform crops anything else. The sleeve is whatever the die-line says. None of these is chosen by looking at the picture; all of them are chosen by looking at where the picture is going.
Every vendor gives you the ratio as a hard control. Midjourney's chart allows up to 14:1 on V8.x and 4:1 for HD, and the same page warns that "aspect ratio isn't the same as image dimensions", which is the sound of a control that fixes the shape and says nothing about the size. OpenAI's sizes are 1024 square, 1536x1024, 1024x1536, or custom multiples of 16 anywhere between 1:3 and 3:1 with no edge over 3840. Google's Flash Lite lists 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9 and 21:9. Different vocabularies, same fact: the shape of the frame is yours to fix.
Once the shape is fixed, the empty band has only a few places it can go. A 3:2 frame can spare its upper third. A 21:9 frame cannot; a third of its height is a strip nobody can set a headline in, so the zone becomes the left third instead. A square spares a lower band. The ratio decides where the words can live, and that is why it is chosen before the subject is placed rather than after the picture comes back.
Ratio limits read 2026-09-18 from docs.midjourney.com (Aspect Ratio, Version), developers.openai.com and ai.google.dev. All three move; the date is part of the fact.
Self-studyWhy the zone is decided before the picture4 min▶
Three things go on every one of Daybreak's six assets that the model never touches: the headline, the origin claim, and the call to action. They are set by hand afterwards, in HTML or SVG, and session 4 owns the reason. What matters tonight is a consequence of that reason: hand-set words need room, and room is a property of the composition, not of the type.
A picture generated edge to edge with the subject in the middle has nowhere for them. The two ways out are both bad. Cropping to make room cuts the subject, so the approved mug loses its rim or its beans. Overlaying the words on the picture puts type on steam and a dark scrim over the thing the client signed off. Both are fixes applied after the fact to a picture that was composed as if it were the finished asset, when it is one layer of a two-layer asset.
So the brief carries a placement instruction that leaves one third quiet, in words: "the upper third is empty negative space", "the mug sits in the right two-thirds, the left third is quiet". Words persuade rather than bind, which is why the reference kit from session 2 includes DB-04, a composition reference whose only content is a subject in position and an empty third. It shows the model the shape the words describe, so the instruction is reinforced by a binding layer instead of left to the sentence alone. Part 3 is what that reference carries and, more importantly, what it does not.
Daybreak's first landing header was a lovely 21:9 with the mug dead centre. The headline had to sit on the steam, so the designer added a dark overlay to make it legible, and the picture the client had approved was gone under a grey wash. The regenerated version, with "the mug in the right two-thirds, the left third quiet" added to the brief and DB-04 in the kit, cost one run. The overlay had cost two days of back and forth about whether the picture was "too dark now".
Subject, light, place: the words the bench counts
6 min liveA composition line is a sentence about where things are. The test of every word in it is whether a painter could be wrong about it. "Lower right" can be wrong. "Oak" can be wrong. "From the left" can be wrong. "Balanced" cannot, because it names nothing, and a word that names nothing gives the model nothing to honour.
LiveA composition line is anchors, not adjectives3 min▶
The bench in session 4 counts concrete anchors in a brief against four lists: colour words (amber, cream, burnt orange), light words (morning light, from the left, soft shadow), placement words (left third, foreground, beside, lower right), and material words (ceramic, oak, paper grain). It also counts numbers and proper nouns. Nothing on any of those lists is an opinion. Every entry names a thing that can be present or absent in the finished picture, which is the same as saying every entry is a thing the painter can get wrong.
That is what makes a composition line written entirely in those words a binding-ish instruction. It is still layer 4, still persuasion, but it is persuasion with a checkable claim per word: either the mug is lower right or it is not, either the light falls from the left or it does not. Ten runs will honour most of those claims most of the time, and the runs that do not are rejected on sight because the miss is visible.
"Balanced", "dynamic", "eye-catching" and "well-composed" are on no list, because they name nothing. A picture cannot fail to be dynamic in a way anyone can point at. They feel like composition words and they are quality words wearing a costume. Strike them and nothing in the instruction is lost, because nothing was there.
Self-studyWhat placement words a model actually honours4 min▶
The vendor documentation promises two things about composition and neither of them is placement. It promises the ratio, as a hard control. It promises reference weight, as a dial. On placement itself, OpenAI's image guide documents difficulty "placing elements precisely in structured or layout-sensitive compositions", and no vendor's parameter list has a control that says "put the subject here". Placement is asked for in words, and words persuade.
The binding layer behind the words is the composition reference. Adobe Firefly names it as a control that matches "outline and depth" with a Strength slider. In a Midjourney kit the same job is done by DB-04, a reference image whose whole content is where the subject sits and which third is empty. The words describe the layout; the reference shows it; the two together are considerably stronger than either.
The honest expectation, then: placement words plus a composition reference land the subject in the right third on most runs, and the runs that miss are rejected on sight and regenerated. That is not a defect to apologise for. It is the same discipline a photographer applies to a contact sheet: shoot the frame, keep the ones where the subject is where it should be. What the discipline needs is a composition line clear enough that a miss is obvious, and that is a property of the words.
| Anchor kind | Examples the bench counts | What it decides |
|---|---|---|
| Colour | amber, burnt orange, cream, ochre, terracotta, black, white | The palette, and whether the picture matches the six others in the set |
| Light | morning light, from the left, side light, soft shadow, backlit, golden hour | Where the shadows fall, which decides where the dark and light areas are and so where type can be read |
| Place | left third, upper third, foreground, beside, behind, on the right, eye level, close-up | Where the subject sits and which third stays empty; the type zone is written in these words |
| Material | ceramic, oak, walnut, linen, kraft paper, glass, brass, paper grain | What the surfaces are made of, which is most of what "looks real" means in a flat style |
| Numbers and names | three beans, one green bean, 3:2, Guji, Daybreak | How many of a thing, and which specific one; a proper noun is a fact the model either knows or does not |
Read 2026-09-18 from docs.midjourney.com, developers.openai.com, ai.google.dev and helpx.adobe.com. Every one of these moves; the date is part of the fact.
A good placement line is short. Daybreak's email hero needs about twenty words: the mug lower right on a worn oak counter, three roasted beans and one green Guji bean beside it, morning light from the left, the upper third empty. Every one of those words is on a list or is a number or a name. Bring the line to session 4 and watch the bench count it.
The reference that carries shape and not colour
4 min liveSession 2 built a kit of four files: one subject reference, two style references and a composition reference, DB-04. The first three say how the picture looks. The fourth says only where things are, and the reason it can sit safely beside the others is that it carries nothing they carry.
LiveOutline and depth, and nothing else2 min▶
Adobe Firefly documents its composition reference as a control that matches "the structure of an image" using "outline and depth", with a Strength slider to say how closely. Read what is not on that list. It does not carry colour. It does not carry line weight, texture or style. That is not a gap in the feature; it is the feature. A reference that carried colour as well as shape would argue with the style references sitting next to it, and the output would be a negotiation between two pictures instead of a layout applied to one look.
In the Midjourney kit the same job is done by DB-04, a file whose only content is the subject's position and the empty third. It is deliberately grey and deliberately dull. The three references beside it, DB-01 for the subject and DB-02 and DB-03 for style at weight 100, own everything about how the picture looks. DB-04 owns where. Four files, no overlap in what each is allowed to say.
Firefly composition reference page updated 2026-06-16, read 2026-09-18 from helpx.adobe.com.
Self-studyHow the six thumbnails become six briefs3 min▶
Each thumbnail from the build-along yields exactly one placement line. The six lines share the subject, the light and the materials, because the picture is the same picture, and they differ in one clause: which third stays empty. "The upper third is empty" for the email hero. "The left third is quiet" for the header. "The lower band is empty" for the three tiles. "The upper band is empty" for the sleeve, where the origin claim will go. Six briefs that differ by a clause are one brief, which is what a set should be.
The reference kit does not change across the six. Same four files, same weights. If the composition reference has to change per ratio, it has been drawn too tightly; a subject in the two-thirds with an empty third beside it is a shape that survives being stretched from 3:2 to 21:9, and a square tile reads it as a lower band without complaint.
The control sheet from session 1 gets one more column: zone. It records which third is empty and lists the words that will fill it, the headline, the claim, the call to action, so that whoever sets the type knows where and what. Those words are on the sheet and never in the brief. Session 4 owns why, and it is not a matter of taste.
Six thumbnails, six placement lines
★ 22 min · everyone drawsOne sheet of paper, six small frames, a pencil. Daybreak's six assets, or six from your own work if you have them. The drawing skill required is the ability to draw a rectangle and an oval, and the rule for the whole exercise: the zone is drawn before the subject, and the line is written in words a painter could be wrong about.
Draw the six frames. Six rectangles at their ratios on one page: a 3:2, a wide 21:9, three squares, and a tall sleeve about 2:3. Do not draw them the same size; draw them the shape they are. The proportions are the first thing a thumbnail teaches you, and a 21:9 drawn as a 3:2 will lie to you about where the words can go.
Place the subject. An oval for the rim and a box for the mug, in the two-thirds of each frame that is not the zone, one thumbnail each. Lower right for the hero, right two-thirds for the header, upper two-thirds for the tiles, lower half for the sleeve. Add three dots for beans and one open circle for the green one. That is the whole drawing.
Dash the zone. A dashed rectangle over the third that stays empty. Inside it, write the words that will go there later: the headline for the hero, the origin claim for the sleeve, the call to action for the tiles. Writing them on the thumbnail is how you find out whether the zone is big enough. It is also the last time they are written anywhere near a brief.
Write the placement line. One sentence per asset, under twenty-five words: where the mug sits, where the light comes from, what the mug stands on, which colours, and which third is empty. Only place, light, material and colour words, plus numbers and names. The six lines should be nearly identical and differ in the last clause.
Check every word. Read each line aloud and strike any word that is not a thing a painter could be wrong about. "Beautiful" goes. "Balanced" goes. "Warm" stays only if it means a colour, and if it does, write the colour instead. What survives is the instruction. Count the survivors; session 4 counts them again on a bench.
Head the page with the picture in one line, the one all six share: a ceramic mug of black coffee on a worn oak counter, steam rising, three roasted beans and one green Guji bean beside it, morning light from the left. Everything below it is placement.
Write the destination and ratio above each frame before drawing anything inside it: "email hero, 3:2", "landing header, 21:9". The ratio comes from the destination, never from the drawing.
Dash the zone first, then place the mug in what is left. If the mug will not fit comfortably in the remaining two-thirds, the zone is too big or the ratio is wrong for a headline of that length; find out now, on paper, not on run eleven.
Write the words the zone will hold inside the dashes, at roughly the size they will be set. A headline that fills the zone edge to edge has no margin; make the zone a little larger or the headline a little shorter, and note which.
Write the placement line under each frame and underline every word that is on a list: colour, light, place, material. Circle every number and name. Anything with neither an underline nor a circle is a candidate for deletion.
Add a zone column to your control sheet and copy in the six zones with the words each will hold. The brief for each asset is the shared picture line plus its placement line. The zone column is for the person who sets the type, and it never travels into the brief.
If you take one thing from tonight. The zone is the composition. Everything else about placing the subject follows from deciding which third stays empty, and that decision is made from the ratio and the destination before anything is generated. A picture composed as if it were the finished asset has to be cropped or covered to become one. A picture composed with its zone empty is finished the moment the type goes on.
Before session 4
25 min- Finish six placement lines for a real asset set from your own work, one per destination, each under twenty-five words, each written only in place, light, material and colour words plus numbers and names. Keep them with the control sheet.
- Find one shipped asset where the headline sits on the subject, from your own work or anyone's, and write the placement line that would have prevented it. Note which third it would have kept empty and what that would have cost the picture. Usually nothing.
- Add a composition reference with an empty third to your kit, or draw one: grey shapes, a subject in the two-thirds, a dashed empty band, no colour and no texture. Name it the way the kit names things and record what it is for.
- Bring a laptop and one prompt you have really used, the longer the better. Session 4 counts its words on a bench, and the count is more useful when the prompt is real and a little embarrassing.
Three questions
Clear all three and session 3 gets stamped in your passport.
1 · Why is the type zone decided before the picture is generated, rather than found in the picture afterwards?
The headline, the claim and the call to action never enter the brief; they are set in HTML or SVG afterwards, and afterwards needs room. Room is a property of the composition, so it is written into the brief as a placement instruction and shown to the model by a reference whose only content is an empty third. Daybreak's dark overlay was what happens when the zone is found instead of decided.
2 · What makes a composition line an instruction rather than a wish?
"Lower right", "from the left", "oak" and "amber" can each be present or absent in the finished picture, which is what makes them checkable. "Dynamic" cannot fail in any way anyone can point at. The bench in session 4 counts the first kind and ignores the second, and the same test works on paper tonight: strike every word a painter could not be wrong about.
3 · What does a composition reference carry into the output, and what does it leave to the other references?
Firefly's own page names the two things matched, outline and depth, and the one dial, Strength. Colour, line and texture are not on the list, which is why DB-04 in the Midjourney kit is drawn grey and dull on purpose. Four files, and no two of them are allowed to say the same thing about the picture.
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 3 cheat sheet · pin this
Six zones are empty and the words for them are sitting in the zone column of your control sheet. Session 4 is why they never go in the brief: the text encoder never saw a letter, three vendors say so in their own documentation, and a bench reads your brief for the eleven characters that would break it. Bring the laptop and the real prompt.