Seedream 5.5 Image Generation

Use dola-seedream-5-5-261028 on ModelSell to generate images, edit references, or decompose an image into layers. All three modes use the same JSON endpoint.

POST https://api.modelsell.com/v1/images/generations
Authorization: Bearer $MODELSELL_API_KEY
Content-Type: application/json

Image generation produces one image per request and supports native RGBA transparency. Layer decomposition can return a base image and up to 99 transparent PNG layers. Layers with regular text can also include rich-text OCR.

Request parameters

Set response_format explicitly. With the API key set to follow the request, ModelSell uses the requested format. With no format set in the request, API key, or channel, ModelSell defaults to b64_json. A fixed API key format takes precedence. The direct Ark API defaults to url. Save images promptly, as upstream URLs normally expire after 24 hours.

This model supports a single output in generation mode. For this model, omit sequential_image_generation, sequential_image_generation_options, stream, and tools. Use layer_decomposition_options.max_images to cap decomposition output.

Sizes and reference images

A custom generation size such as 2048x1152 must have 921,600–4,624,220 total pixels and an aspect ratio from 1:16 to 16:1. With a preset, describe your desired aspect ratio in the prompt. In decomposition, the base image and each layer retain their respective aspect ratios from the source.

Each reference must have an aspect ratio from 1:16 to 16:1 and be at most 30 MB. Use a URL the upstream provider can reach, or a string such as data:image/png;base64,BASE64_DATA with a lowercase media subtype. The ModelSell gateway caps the request body at 128 MiB. URLs can keep requests with many large references within that limit.

Generate and save a transparent PNG

curl --fail-with-body -sS https://api.modelsell.com/v1/images/generations \
  -H "Authorization: Bearer $MODELSELL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "dola-seedream-5-5-261028",
    "prompt": "One isolated red rain boot with a crisp outline on a completely transparent background",
    "size": "1K",
    "background": "transparent",
    "output_format": "png",
    "response_format": "b64_json",
    "watermark": false
  }' > transparent.json

Decode the returned image bytes to a PNG file.

import base64
import json
from pathlib import Path

data = json.loads(Path("transparent.json").read_text())["data"][0]
Path("transparent.png").write_bytes(base64.b64decode(data["b64_json"]))

Edit a reference image

Send this JSON to the same endpoint. Replace image with the URL of the image you want to edit. For multiple references, pass an array and describe each image in order in the prompt. An ordinary opaque input can also produce a transparent PNG.

{
  "model": "dola-seedream-5-5-261028",
  "prompt": "Add round sunglasses to the subject while preserving the pose and composition",
  "image": "https://arkdocs-en.tos-ap-southeast-1.volces.com/images/image-generation/seedream-5-5-multi-round-editing-input.png",
  "size": "2K",
  "output_format": "png",
  "background": "opaque",
  "response_format": "url"
}

For a subsequent edit, pass the previous result as image and describe what to change and preserve in the new prompt.

Decompose layers and recognize text

This example uses an Ark sample poster and requests up to 15 images including the base image.

curl --fail-with-body -sS https://api.modelsell.com/v1/images/generations \
  -H "Authorization: Bearer $MODELSELL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "dola-seedream-5-5-261028",
    "prompt": "Extract the title and signature as separate editable text layers. Preserve the original text.",
    "image": "https://arkdocs-en.tos-ap-southeast-1.volces.com/images/image-generation/ocr-test2.png",
    "layer_decomposition": true,
    "layer_decomposition_options": {
      "max_images": 15,
      "text_recognition": true
    },
    "size": "auto",
    "output_format": "png",
    "response_format": "url",
    "watermark": false
  }' > layers.json

Set response_format to b64_json to embed the base image and each layer separately. To decompose without OCR, omit text_recognition or set it to false.

Response fields

Both bounding_box.absolute and bounding_box.normalized contain [left, top, right, bottom]. Absolute coordinates measure pixels from the top-left corner of the base image. Normalized coordinates range from 0 to 1000 on each axis. To reassemble the image, scale each layer to its bounding box and stack the layers by ascending z_index.

text_recognition.text contains the layer's full recognized text, with line breaks between lines. Each lines[] item contains that line's bounding_box and runs[]. Each run contains text and style. Style fields include similar_font_family (unknown when unrecognized), foreground_color (#RRGGBB), and the booleans bold, italic, underline, strikethrough, shadow, and translucent. The model estimates style. In a ModelSell production sample, the model read both lines correctly and identified bold and italic text. It missed one visible underline.

This excerpt shows one data[] item from a production text layer. The Base64 value is a placeholder.

{
  "z_index": 1,
  "name": "Bold navy title text “SUNNY DAY”",
  "description": "Bold navy title text “SUNNY DAY”",
  "size": "1677x625",
  "output_format": "png",
  "background": "transparent",
  "b64_json": "BASE64_PNG_DATA",
  "bounding_box": {
    "absolute": [43, 135, 1162, 552],
    "normalized": [36, 155, 964, 634]
  },
  "text_recognition": {
    "text": "SUNNY DAY",
    "lines": [
      {
        "bounding_box": {
          "absolute": [193, 272, 1015, 434],
          "normalized": [160, 313, 842, 499]
        },
        "runs": [
          {
            "text": "SUNNY DAY",
            "style": {
              "similar_font_family": "Arial",
              "foreground_color": "#111E3E",
              "bold": true,
              "italic": false,
              "underline": false,
              "strikethrough": false,
              "shadow": false,
              "translucent": false
            }
          }
        ]
      }
    ]
  }
}

References

• BytePlus Image Generation API
• Volcengine Ark Image Generation API
• Official Dola Seedream 5.5 tutorial
• ModelSell image generation API reference