Answer: Seedream 5.5 renders editorial images at 1K, 1.5K or 2K in eleven ratios — including 21:9 and 2:1 for covers and banners — and its layer mode can isolate lettering as its own layer with text recognition on. The practical workflow is: render the picture with no text, place type in design software, and use layers when the layout will be revised.
Generated type is good and typeset type is better. The reason to separate them is not quality on the first render; it is the fifth revision, when a headline changes and the artwork does not.
Ratios for print-adjacent work
| Format | Ratio | Size to render |
|---|---|---|
| Magazine cover | 3:4 or 2:3 | 2K |
| Wide banner | 21:9 | 2K |
| Skyscraper unit | 1:2 | 2K |
| Square campaign tile | 1:1 | 1.5K |
| Spread | 3:2 | 2K |
1.5K costs the same 2 credits as 1K, so the only decision is whether the final needs 2K — for anything printed, yes.
The picture/type split
A layout that survives review has three layers of responsibility:
- The picture — generated, no lettering. Put "no readable text" in the prompt; the model is good enough at type to invent it otherwise.
- The type — set in design software, where a change is a text edit.
- The composite — exported once, then checked at 100%.
If the house style is to let the model draw the headline, do it at 2K and budget one re-render per copy change. That is a legitimate choice for one-off editorial; it is a bad choice for a monthly cover.
Using layer decomposition in an editorial flow
- Split the finished image into 8–15 layers at 2K with text recognition on.
- The background layer becomes reusable artwork for the next issue.
- The lettering layer tells you where type sits, which is useful even if you replace the letters.
- Keep the stack with the issue's source file, so the next revision edits parts rather than re-rendering everything.
Cost note: a layer split settles on usage and lands above a flat 2K render (4 credits), because it delivers many outputs. Budget it for artwork that will be reused, not for every image.
Prompt habits for editorial
| Goal | Say this |
|---|---|
| Room for a headline | "generous negative space across the upper third" |
| No invented lettering | "no readable text, no signage, no logos" |
| Print-safe palette | "flat, print-friendly colours; avoid gradients" |
| Believable perspective | "crisp focal hierarchy, believable perspective" |
Those four clauses appear in the published examples, and they are the differences between a picture that accepts type and one that fights it.
What a cover costs
| Job | Credits |
|---|---|
| 1.5K draft, no text | 2 |
| 2K final, no text | 4 |
| 2K final + layer split (12 layers) | usage, above 4 |
| Ten cover concepts at 1.5K | 20 ($0.20) |
Ten concepts for twenty cents is the real argument: concept the covers, pick two, finish those at 2K, split only the winner.
Where this goes wrong
- Asking for the headline in the prompt and tweaking it five times. Each tweak is a ~4-credit render and a fresh set of compromises.
- Rendering a 21:9 banner with no negative space. The type ends up over the subject.
- Splitting every image. Layers are for reuse, not for habit.
- Expecting 4K. The model stops at 2K; scale in design software if needed.
FAQ
Is the model good at text? Yes — good enough that you must tell it "no readable text" when you plan to set type yourself.
Can I get transparent text layers? Layer outputs are images at a chosen size; treat the lettering layer as a guide.
Which ratio for a Story cover? 9:16.
Does text recognition cost extra? It is part of the usage-based settlement, inside the ceiling.
The ratio set and the layer options are on Callirra's Seedream 5.5 page; the published examples there are worth reading for the "no readable text" habit alone.