Seedream 5.5

Use model doubao-seedream-5-5-261028 with POST https://apimaster.ai/v1/images/generations, a JSON body and Authorization: Bearer YOUR_API_KEY. Create your APIMaster key in the console.

Supports text-to-image, editing with up to ten references, successive editing, native transparent RGBA output, layer decomposition and rich-text OCR. Normal generation returns one image. Decomposition returns a base and separate layers. Streaming, sequential image generation and web search are not supported. Use the synchronous endpoint; allow at least 300 seconds before a client timeout and avoid blind resubmission.

Official base pricing

Checked against Volcengine pricing on November 13, 2026. The configured conversion was CNY 7 per USD.

Item	CNY per image	Approximate USD base price
Every input reference, including the first	0.05	0.00714286
Output with width × height ≤ 2,610,000 pixels	0.13	0.01857143
Output above 2,610,000 pixels	0.23	0.03285714

Generation and decomposition use the same output rates. Charge each base/layer separately using its actual data[].size; the longest edge does not determine the tier. 2048x1024 is in the lower tier. Two inputs plus one high-tier output cost CNY 0.33 before account/channel multipliers. Unlike Seedream 5.0 Pro, the first reference is not free. Actual account charges include applicable channel/user group coefficients and integer quota rounding. See the live model card and usage logs.

Text-to-image

curl --max-time 300 "https://apimaster.ai/v1/images/generations" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model":"doubao-seedream-5-5-261028",
    "prompt":"A cream landscape poster, a red circle, a blue square, heading HELLO SEEDREAM",
    "size":"1.5K",
    "output_format":"png",
    "background":"opaque",
    "response_format":"url",
    "watermark":false
  }'

Parameters

Field	Type / default	Rules
model	required string	doubao-seedream-5-5-261028.
prompt	string	Required for generation; optional for decomposition. Describe content, style, aspect ratio, text or the editing location. Official writing guidance recommends ≤300 Chinese characters or ≤600 English words; this is guidance rather than a byte limit.
image	string or string[]	Public URL or data:image/png;base64,.... Up to ten generation references; exactly one input for decomposition. Local filesystem paths do not work.
size	string	Generation defaults to 2K; 1K, 1.5K, 2K or exact WIDTHxHEIGHT. Decomposition defaults to auto, accepts only auto and these three tiers.
output_format	string, png	png or jpeg. For decomposition this controls only the base; layers are always PNG.
background	string, auto	auto, transparent or opaque, for generation only. Transparent output requires PNG; omit this field entirely for decomposition.
response_format	string, url	url or b64_json. Read the matching field in every data[] item.
watermark	boolean, true	Set false to omit the visible AI-generation watermark.
optimize_prompt_options	object	Only {"mode":"standard"} is supported by 5.5; fast is not supported.
layer_decomposition	boolean, false	Enable decomposition of one PNG/JPEG image.
layer_decomposition_options	object	Decomposition options below.
layer_decomposition_options.max_images	integer, 30	2–100, including the base. A maximum, not a guaranteed count.
layer_decomposition_options.text_recognition	boolean, false	Request rich-text OCR for regular text layers.
safety_identifier	string	Stable English user identifier, at most 64 characters; use a hash instead of personal contact information.

Do not use another model's resolution, quality, n or seed to control this model. These are not listed as its corresponding generation controls in the official API. Set decomposition limits with layer_decomposition_options.max_images.

Sizes and input restrictions

For ordinary generation, choose a tier (1K, 1.5K, 2K) and describe the aspect ratio in the prompt, or supply exact dimensions. Exact output area must be 921,600–4,624,220 pixels, with aspect ratio 1:16–16:1, satisfying both constraints. 2880x1440 is valid; 512x512 is too small. There is no 4K tier.

Decomposition does not accept exact dimensions. With auto, base/layer sizes within the supported range retain their original dimensions where possible; smaller/larger items scale toward 1K/2K while preserving their aspect ratio. Returned layers can occupy different billing tiers.

Input constraint	Generation/editing	Decomposition
Count	Up to 10	Exactly 1
Formats	JPEG, PNG, WebP, BMP, TIFF, GIF, HEIC, HEIF	PNG, JPEG
File size per image	≤30 MB	≤30 MB
Pixel area	196–36,000,000; both sides >14 px	262,144–36,000,000
Aspect ratio	1:16–16:1	1:16–16:1

Use lowercase MIME types in data URLs. Remote input URLs must be accessible to the service.

Editing and transparent images

{
  "model":"doubao-seedream-5-5-261028",
  "prompt":"Keep the layout of the first image; recolor only its circle using the second reference",
  "image":["https://example.com/source.png","https://example.com/reference.png"],
  "size":"1.5K",
  "output_format":"png",
  "watermark":false
}

For one input, image can be a string. For successive edits, explicitly pass the previous output as the next input: this is not a stateful conversation, and each request is billed separately. Describe coordinates/boxes/arrows in the prompt or draw marks on the reference; there is no separate documented mask positioning parameter for this interface.

For native transparent generation, omit image if unnecessary and set background: "transparent", output_format: "png", with a prompt such as “a translucent blue glass butterfly, isolated asset, transparent background”. Editing can use up to ten references; they need not already have alpha channels. transparent plus jpeg is invalid. A response's background describes the actual transparency (transparent or opaque).

Decomposition and rich-text OCR

{
  "model":"doubao-seedream-5-5-261028",
  "image":"https://example.com/poster.png",
  "prompt":"Separate the heading, circle and square into editable layers",
  "layer_decomposition":true,
  "layer_decomposition_options":{"max_images":4,"text_recognition":true},
  "size":"auto",
  "output_format":"png",
  "watermark":false
}

Supports one base plus up to 99 layers. max_images: 4 includes the base. Omit background; every layer is transparent PNG. An individual layer failure makes the whole request fail rather than returning partial success.

Read all data[] entries:

• url/b64_json, size and output_format identify the artifact and its actual dimensions; background is available where applicable.
• z_index: 0 identifies the base; larger indices appear above smaller ones.
• name and description describe the layer.
• bounding_box.absolute is [left,top,right,bottom] in output-base pixels. Resize a layer to (right-left) × (bottom-top) and place it at (left,top); layer file dimensions can differ from its canvas footprint.
• bounding_box.normalized uses 0–1000 coordinates. Multiply horizontal/vertical coordinates by target canvas width/height and divide by 1000.
• text_recognition appears only when requested and the layer contains recognized regular text. Handwriting, artistic lettering and non-text layers may omit it.

OCR contains text and lines[]; each line has its own bounding_box and styled runs[]. Each run has text and style, which can contain similar_font_family, foreground_color (#RRGGBB), bold, italic, underline, strikethrough, shadow and translucent. These are approximate visual interpretations, not exact original design properties; font recognition can return unknown.

Response, storage and failures

The response contains created, model, data[], usage.input_images and usage.generated_images. The generated count includes base and layers. Preserve the complete response metadata. Original provider download URLs expire after 24 hours; promptly download/archive your assets.

Correct invalid dimensions, reference counts or incompatible transparency options before retrying. Handle authentication, permissions, moderation, rate limits and upstream failures separately; never interpret an error as a successful generation. IPM limits count generated images per account/model version; decomposition reserves max_images and returns unused capacity after completion, while billing uses actual produced images.

Official sources: Seedream 5.5, Image generation API, pricing. Parameters checked November 13, 2026.