One good Midjourney poster takes about ten minutes. Fifty that look like they belong together takes a decision made before the first one, and most sellers make it after the twentieth.
The failure is easy to spot in a shop. Scroll the listings and the prints are individually fine, but the palette shifts, the rendering style wanders between painterly and photographic, and the ninth poster has a sun flare that nothing else has. Each image was judged on its own and passed. The collection was never judged at all.
That matters commercially, not just aesthetically. A buyer who likes one of your posters is the most likely person on earth to buy a second, and a coherent range is what converts that. An incoherent one gives them nothing to buy next.
Style references beat style words
The instinct is to describe the style in words and reuse the description. It half works, and the half that fails is the half you notice.
The problem is that descriptive style words are interpreted relative to the subject. "Muted tones" against a Norwegian fjord produces greys and blues. The same phrase against a Moroccan desert produces sand and ochre. Both are muted; neither matches the other. You wrote one instruction and got two palettes, because the model resolved your adjective against a different scene each time.
A style reference works differently. --sref points at an actual image and asks the model to carry its qualities across. Because the target is fixed rather than described, it does not drift when the subject changes.
The practical method:
- Generate until you have one poster whose look you would happily repeat fifty times.
- Take its URL, or its
--srefcode, and hold it. - Every subsequent generation in the range uses that same reference, unchanged.
- When you want a second collection, you start a second reference — you do not edit the first.
The discipline is in step three. The temptation, around image fifteen, is to tweak the reference because a particular place is not working. Do not. You will have two styles and no way back to the first.
Seeds are not what you think
Locking --seed feels like it should hold a series together. It does not, and the misunderstanding wastes a lot of time.
A seed fixes the random noise the image starts from. With an identical prompt it reproduces an identical image, which is genuinely useful for reproducing a result you have lost. Change the prompt — a new city, a new landscape — and the noise no longer maps to anything comparable. You get a different image that happened to begin from the same starting point, which is not the same as a consistent one.
Seeds are for reproduction, not consistency. Style references and a fixed prompt skeleton are for consistency.
Fix the skeleton, vary one slot
Write your prompt so that exactly one part changes between posters.
[SUBJECT], [fixed landscape treatment], [fixed lighting],
[fixed rendering style] --ar 1:1 --sref [held reference]Everything except [SUBJECT] is typed once and never touched. That sounds obvious written down, and almost nobody does it — prompts get edited in place, adjectives accumulate, and by poster thirty the instruction bears no resemblance to poster one.
Keep the skeleton in a text file rather than in your prompt history. History is a log of what you tried, not a record of what you decided.
Where consistency always breaks
Three places, reliably.
Places the model knows too well. Paris, Tokyo, New York and Santorini carry so much training weight that they override style instructions. Ask for a muted minimal Paris and you will fight the Eiffel Tower and a warm golden palette that arrives whether you asked or not. Famous places pull toward their own clichés. Expect to spend more attempts on the landmarks than on anywhere else, and consider whether your range needs them at all — they are also the most competitive listings on Etsy.
Places the model barely knows. The inverse problem, and the more damaging one, because it fails quietly. A small town produces plausible generic scenery that is not that town. Buyers from the region notice immediately, and that is the review that hurts. For lesser-known places, prompt the terrain rather than the name: the coastline, the building materials, the vegetation, the light. Let the poster's text do the naming.
Vertical subjects in horizontal frames. A tall lighthouse, a cathedral spire and a canyon do not compose the same way as a broad coastline. If your range mixes them, the subject sits differently in the frame each time and the collection loses its rhythm even when the palette is perfect. Either commit to one compositional family, or mount every render on a fixed layout so the frame is constant even when the image is not.
Generate square, mount later
The single most useful structural decision: generate at 1:1 and mount the artwork on a taller sheet with a text band beneath it, rather than generating directly at 2:3 or 3:4.
Three reasons.
The composition is easier to control. Square framing avoids the model's habit of adding foreground filler to justify a tall canvas — filler that differs every time and quietly breaks your consistency.
You keep the resolution. A square render mounted on a 3:4 sheet uses its full width, and the band beneath is generated by your layout rather than by the model. Nothing is upscaled to fill space.
And you can change print ratio later without regenerating anything. The same square artwork mounts on 2:3, 3:4, 4:5 or A-series. If you generated at 2:3 and later want A4, you are re-rendering the range.
The trade-off is that you need a layout that does the mounting reliably, with margins and typography that hold across every poster. That is a one-off setup, and it is the part that makes a range look designed rather than generated.
Judge the grid, never the image
The habit worth building: never approve a poster on its own.
Put every render into one contact sheet — a grid of thumbnails, all at the same size — and look at that. Consistency problems are invisible at full size and obvious at thumbnail. The one image that is slightly warmer, slightly sharper, slightly more saturated leaps out of a grid and hides completely on its own.
This is also, not coincidentally, how a buyer sees your shop: as a grid of thumbnails. If your range does not hold together at that size, it does not hold together anywhere that matters.
The order that works
- Decide the print ratio and the layout before generating anything.
- Generate candidates until one is worth repeating.
- Lock it as your style reference. Write the prompt skeleton to a file.
- Generate the range, changing only the subject slot.
- Review as a grid, not as images.
- Regenerate the outliers against the same reference — never by editing it.
The expensive version of this is discovering at poster forty that the first fifteen do not match, and choosing between regenerating them and shipping a shop that looks like three shops.
Frequently asked questions
What is the most reliable way to keep Midjourney consistent across a series?
A style reference image with --sref, reused unchanged across every generation in the range. It holds palette, rendering and mood far more tightly than descriptive words do, because you are pointing at an example rather than describing one. Descriptive style words drift as the subject changes — 'muted' reads differently against a desert than against a fjord — while a reference gives the model a fixed target that does not move with the subject.
Should I lock the seed to keep a poster series consistent?
No, and this is the commonest mistake. A locked seed with a changed prompt does not give you the same style with a new subject — it gives you a different image that happens to start from the same noise. Seeds are for reproducing one specific result, not for holding a look across many. Style references and a fixed prompt skeleton are what actually carry a range.
How many posters should be in an Etsy poster series?
Enough that a buyer viewing your shop sees a collection rather than a handful of unrelated prints, which in practice means about twelve to twenty in one visual style. Below roughly eight it reads as experiments. The number matters less than the discipline: twelve posters in one coherent style outsell forty in five styles, because the second shop looks like a stock feed and the first looks like a designer.
Why do some places in my series look wrong even though the prompt is the same?
Because the model knows some places far better than others. Paris, Kyoto and New York carry enormous training weight and will dominate your style instruction; a small coastal town has almost none and the model fills the gap with generic scenery. The fix is to prompt the landscape rather than the place name for lesser-known locations — describe the terrain, the light and the architecture, and let the title do the naming.
Does the aspect ratio need to be the same across a poster series?
The generated ratio does not, but the printed one does. A shop mixing 2:3 and 4:5 prints looks unplanned on a wall and complicates your listings, since each ratio needs its own print sizes. Generate square if you intend to mount the artwork on a taller sheet with a text band, or generate directly at the print ratio if the image runs full bleed. Decide once, before the first fifty.