Seedream 5.5 Layer Decomposition

Seedream 5.5 Layer Decomposition (Image-to-Image)

What is Seedream 5.5 Layer Decomposition?

Seedream 5.5 Layer Decomposition is ByteDance's image-to-image model that takes one finished image and splits it into a clean base layer plus a stack of named, transparent PNG layers — backgrounds, objects, text and semi-transparent elements — with no manual masking. Every returned layer carries a name, a short description and a bounding box (absolute pixels and normalized 0–1000), so a single flat picture becomes a set of editable, re-composable assets.

Key Features

• •Automatic separation into up to 100 layers, base image included, or fewer when you cap max_images.
• •Native RGBA transparency: each element returns as a transparent PNG with clean alpha edges, ready to recomposite.
• •Named, positioned layers: every layer comes back with a name, a description and a bounding box.
• •Optional text recognition: layers classified as formatted text can return their text with font family, size and colour.
• •Prompt-guided or automatic: leave the prompt empty to find every major element, or name elements to target specific layers.
• •Resolution control: auto to follow the input, or 1K, 1.5K and 2K.

Best Use Cases

• •Game and motion design: turn a flat illustration into reusable 2.5D layers for parallax and depth.
• •Marketing and re-layout: pull a poster apart into background, product and text layers, then rebuild it for a new aspect ratio.
• •E-commerce: isolate products and backgrounds for colour, material and scene variations.
• •Design hand-off: deliver editable transparent assets instead of one flattened render.

Prompt Tips and Output Quality

Feed a clean, well-lit composite with distinct, separable elements and space around each object; crowded or low-contrast scenes separate less cleanly. Leave the prompt empty for a full automatic split, or name the elements (for example, "the locomotive, the tunnel, the blocks") to control which layers return. In testing, one toy-train scene returned a clean inpainted base plus seven crisp, correctly named transparent layers — including see-through tunnel interiors and gaps between letter blocks — with no artifacts.

Using Seedream 5.5 Layer Decomposition

Seedream 5.5 Layer Decomposition turns one finished image into a clean base layer plus named, transparent PNG layers you can re-edit and recompose. Send exactly one image; the model returns the base (z-index 0) and each element as its own transparent PNG with a name, description and bounding box.

How to get clean layers

Start from a composite with clear, separable elements and some space around each one — distinct objects on a readable background separate best. Then decide how the split is driven:

• •Leave the prompt empty to let the model find and name every major element.
• •Name elements in the prompt ("the red train, the tunnel, the letter blocks") to target specific layers.

Choosing parameters

• •max_images caps how many images come back, base included (2–100). Lower values give a focused set; higher values split more finely.
• •size sets base and layer resolution: auto follows the input, or force 1K, 1.5K or 2K.
• •text_recognition returns structured text (font family, size, colour) on layers classified as formatted text — turn it on for posters, labels and localization.
• •output_format sets the base image format only; layers are always transparent PNG.

Recommended Settings

• •Clean demo or reusable assets: prompt empty, max_images 10, size auto, text_recognition off — tested on a toy-train scene, returning a clean base plus seven correctly named transparent layers.
• •Fine-grained breakdown: raise max_images toward 30 for busier compositions.
• •Posters and localization: turn text_recognition on and name the text blocks you want returned.
• •Crisp detail: pick size 2K; use auto to keep the input's resolution.

Each returned image — the base and every layer — is produced at the resolution you choose, so keep max_images modest when you only need the main elements.

Examples to try

Each example below ran end to end on this page at the settings shown; every one starts from a single flat image and returns a clean base layer (z-index 0) plus named, transparent PNG layers.

• •Automatic decomposition (the pre-filled example). Prompt: left empty. A wooden toy-train diorama fans out into a clean inpainted base plus seven named transparent layers — the train, tunnel, house, trees and letter blocks. Tip: leave the prompt empty and keep max_images modest to get just the main elements, not an over-fine split.
• •Prompt-guided targeting. Prompt: "Isolate the vinyl record, its printed center label, the album sleeve and the turntable platter." Naming four elements returned exactly those four as separate transparent PNGs, each with a name and a bounding box. Tip: list only the pieces you need and the model targets them, leaving the rest in the base.
• •Native RGBA transparency. Prompt: left empty. A translucent moon jellyfish lifts off the dark water as one RGBA layer, its bell see-through and tentacles kept soft-edged. Tip: transparent output is always PNG — recomposite the layer straight onto a new background with no manual masking.

More to try: "Separate the product, its shadow and the backdrop", "Split this poster into background, headline and logo layers", "Return each character and prop as its own transparent layer".

Does every layer come back transparent?

Yes — layers are always transparent PNG; only the base image honours output_format.

Can I choose which elements become layers?

Yes — name them in the prompt, or leave it empty to auto-detect every major element.

Does it extract editable text?

On layers it classifies as formatted text it can return the text with font, size and colour; not every text layer qualifies.

How many layers can it return?

Up to 100 including the base, capped by max_images (default 30).

Do I need to remove the background first?

No — the base layer is returned with the foreground elements already separated out.