Seedream 5.5 Layer Decomposition: Rebuild and Edit a Design

Seedream 5.5 layer decomposition returns a base image, transparent PNG layers and placement metadata. To turn those outputs into an editable design, your application must resize and position each layer, preserve stacking order, and create text objects from any available OCR. Downloading the PNGs is the start of that work.

Decide what should remain editable

Start with one approved poster and one concrete change: move a product, replace a heading or create another language version. The output you need depends on that change.

Request the parts your workflow needs

curl -X POST https://api.evolink.ai/v1/images/generations \
  -H "Authorization: Bearer $EVOLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "doubao-seedream-5.5-layerize",
    "image_urls": ["YOUR_PUBLIC_IMAGE_URL"],
    "quality": "auto",
    "output_format": "png",
    "n": 30,
    "model_params": {"text_recognition": true}
  }'

Reassemble Seedream 5.5 PNG layers

{
  "base": "base.png",
  "layers": [
    {"file": "layer-1.png", "box": [120, 80, 420, 280], "z": 1}
  ]
}

import json
import sys
from pathlib import Path
from PIL import Image

manifest = Path(sys.argv[1])
data = json.loads(manifest.read_text())
root = manifest.parent
with Image.open(root / data["base"]) as image:
    canvas = image.convert("RGBA")
width, height = canvas.size
layers = data["layers"]
orders = [item["z"] for item in layers]
if any(type(z) is not int or z <= 0 for z in orders):
    raise ValueError("Layer order must be a positive integer")
if len(set(orders)) != len(orders):
    raise ValueError("Duplicate layer order")
for item in sorted(layers, key=lambda layer: layer["z"]):
    box = item["box"]
    if len(box) != 4 or any(type(v) is not int for v in box):
        raise ValueError("Expected four integer box coordinates")
    left, top, right, bottom = box
    if not (0 <= left < right <= width and 0 <= top < bottom <= height):
        raise ValueError("Box outside the base image or empty")
    with Image.open(root / item["file"]) as image:
        layer = image.convert("RGBA").resize(
            (right - left, bottom - top), Image.Resampling.LANCZOS
        )
    canvas.alpha_composite(layer, dest=(left, top))
canvas.save(root / "reassembled.png")

Turn OCR into editable text

When replacing a text layer with native text, hide its raster counterpart so the old and new letters do not overlap. If one raster layer contains several lines and you replace only one, mask or remove just that line while preserving the others. Alternatively, rebuild all text you need before hiding the whole raster layer. Save text objects, font choices and layout in your application document. OCR data by itself is not a saved editable design.

Adapt the design without distorting it

A square poster converted to a tall ad needs more than coordinate multiplication. Choose whether to fit the old design with margins, crop it, or rearrange its elements. Independent horizontal and vertical scaling can stretch products and type. Moving an object can also reveal imperfect reconstructed background content; inspect the newly exposed area instead of assuming it is the original hidden scene.

For a language variant, replace one heading first. Check glyph coverage, line breaks, text-box height, visual hierarchy and overlap with foreground objects. Keep source and edited states separate so a reviewer can compare them. Translation and reflow are application work, not an automatic consequence of receiving OCR.

Review the result before making more variants

Review the reconstructed source before changing it, then inspect the edited result at its delivery size. Check product identity, edges, background fill, occlusion, wording and final export. Coordinate-correct placement does not establish pixel-perfect reconstruction or acceptable design quality.

Frequently asked questions

Does Seedream 5.5 return a PSD?

The documented EvoLink route returns images and metadata, not a native PSD. An editor or exporter must build the design document.

Does n: 30 mean 30 separate foreground layers?

No. It caps all output images, including the base. The actual count may be lower; read the returned items.

Should I use the layer PNG's original size?

Use its base-image bounding box for placement. The file can be larger than that rectangle, so resize it before compositing.

Can I rebuild from normalized coordinates?

Yes, convert each corner from the 0–1000 coordinate range using the target canvas dimensions. Changing aspect ratio still requires an explicit fit, crop or reflow choice.

Will every text layer contain OCR?

Does OCR identify the exact original font?

Should I regenerate when layers are misaligned?

First check stacking, box interpretation and scaling. Those composition mistakes can be corrected locally. Missing or unusable visual content may need editing or another decomposition.

Do these examples prove EvoLink output quality?

No. They explain the documented workflow using an official sample and local file operations. They do not measure API success rate, reconstruction fidelity or speed.

Sources

• BytePlus Seedream 5.5 guide: official layer/OCR capabilities and sample package; provider field names differ from EvoLink.
• EvoLink Layerize API documentation: model ID, inputs, limits, task workflow and settlement.
• Pillow Image reference: local resizing and alpha compositing.