Answer: Seedream 5.5's layer decomposition mode takes one image and returns it as a stack of separate layers — between 2 and 100, default 30 — at 1K, 1.5K or 2K, with optional text recognition that turns lettering into its own layer. It accepts exactly one reference image; the output count, size and text handling are the three settings.
This is the least common feature in the image models on this platform, and the one that changes what a rendered image is for. A flat PNG is a picture; a layer stack is a working file.
The three settings
| Setting | Options | What it decides |
|---|---|---|
| Layers | 2–100, default 30 | how finely the image is separated |
| Size | 1K · 1.5K · 2K | resolution of the returned layers |
| Text recognition | on / off | whether lettering becomes its own layer |
Everything else is inherited from the source image: aspect, palette, content.
What it is good for
Poster revisions without a repaint. The headline becomes a layer, so swapping copy is a text edit rather than a new generation — and the type stops being a picture of type.
Product isolates. A still life separates into product, surface and background, which is the difference between an asset and a background you can reuse.
Localisation. Text on its own layer means a translated layout does not need a new render.
Style transfer with control. Re-light or re-grade a background layer while the subject stays untouched.
What it is not
- Not a segmentation model. Edges follow the model's judgement, not a matte; hair and glass will not be pixel-perfect.
- Not a vectoriser. You get raster layers, not paths.
- Not free. Layer splits settle on usage with a credit ceiling, because the charge covers the references and every delivered output.
How to ask for it
The mode takes one image and no prompt-driven composition — the brief is the settings plus, at most, a short instruction about what should stay whole:
Source: @Image1
Layers: 12
Size: 2K
Text recognition: on
Keep the product and its shadow as one layer.The last line is the useful habit: telling the model what not to split is how you avoid a product cut into twelve fragments.
Costs, roughly
| Job | Credits |
|---|---|
| 1K, 6 layers | low end of usage, near a normal edit |
| 2K, 12 layers | above a flat 2K image (4 credits) |
| 2K, 30 layers with text | the expensive end — check the quote |
The pattern to remember: the flat per-image price (2 credits at 1K/1.5K, 4 at 2K) is for one output. Layers are many outputs, so they are billed on usage, capped.
A workflow that uses it well
- Generate the image at 1.5K. Two credits, same price as 1K.
- Decide the asset's future — is it a final flat image, or will it be revised?
- If it will be revised, split it at 2K with 8–15 layers and text recognition on.
- Keep the layer stack with the source so the next revision edits copy, not pixels.
Where it goes wrong
- Asking for 100 layers on a simple image. You get fragments, not structure. Ten to twenty is usually the useful range.
- Expecting transparency. Layers are stacked; the background layer carries the background.
- Splitting a low-detail image. At 1K, letters and fine edges separate poorly — split at 1.5K or 2K.
FAQ
Can I control which parts split? Only indirectly, by the layer count and a line about what to keep whole.
Does text recognition fix typos? It isolates text as a layer; spelling is still a review step.
Can I split a photo I did not generate? Yes — one reference image, from your library.
Are layers delivered as separate files? Yes, as a set of images at the chosen size.
The mode's settings are on Callirra's Seedream 5.5 page, and the twenty-edit study there shows a single base image revised repeatedly — layer decomposition is the version of that workflow that survives into production.