Seedream 5.0 Image Generation

Complete Seedream 5.0 Pro and Flash parameters with text, reference-image, interactive-editing, layer-decomposition, and transparent-background examples.

Use the ModelSell OpenAI Images-compatible endpoint for the Seedream 5.0 family:

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

Supported Models

Model	Model ID	Capabilities
Seedream 5.0 Pro	`doubao-seedream-5-0-pro-260628`	Text-to-single-image, one or multiple references, interactive editing, layer decomposition, and transparent backgrounds; up to 10 references; 1K/1.5K/2K; prompt optimization `standard` and `fast`
Seedream 5.0 Flash	`doubao-seedream-5-0-flash-260915`	Same generation, interactive-editing, layer-decomposition, and transparent-background capabilities as Pro with faster generation; up to 10 references; 1K/1.5K/2K; prompt optimization `standard` only

Both models return a single image per request (except for layer decomposition) and do not support image sequences, web search, or streaming. For latency-sensitive workloads, use Flash or Pro's fast prompt-optimization mode.

All Request Parameters

Field	Type	Required	Default	Description
`model`	string	Yes	-	A model ID from the table above
`prompt`	string	Yes	-	Prompt in Chinese, English, or other supported languages such as Japanese, Korean, French, German, and Spanish; keep it within about 300 Chinese characters or 600 English words
`image`	string / string[]	No	-	Up to 10 public URLs or `data:image/<format>;base64,<data>` values
`size`	string	No	`2K` (`auto` for layer decomposition)	Resolution tier or `widthxheight`; do not mix the two
`layer_decomposition`	boolean	No	`false`	Splits `image` into a base image plus transparent PNG layers; requires `image`
`background`	string	No	`opaque`	`transparent` or `opaque`; `transparent` requires exactly one input image with an alpha channel and outputs png
`optimize_prompt_options.mode`	string	No	`standard`	`standard` or `fast`; Flash currently supports only `standard`
`output_format`	string	No	`jpeg`	`png` or `jpeg`
`response_format`	string	No	`url`	`url` or `b64_json`; URLs expire after 24 hours
`watermark`	boolean	No	`true`	Adds or removes the AI-generated watermark

Any other native Volcengine Ark image generation parameter not listed here is also forwarded to the upstream unchanged.

For image generation, both models accept 1K, 1.5K, and 2K (default 2K), or custom dimensions between 921,600 and 4,624,220 total pixels with an aspect ratio from 1:16 to 16:1. 1.5K costs the same as 1K. Layer decomposition accepts only 1K, 1.5K, 2K, or auto (default). Each reference must be directly reachable, no larger than 30 MB, and use jpeg, png, webp, bmp, tiff, gif, heic, or heif.

Text To Image

curl https://api.modelsell.com/v1/images/generations \
  -H "Authorization: Bearer $MODELSELL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "doubao-seedream-5-0-pro-260628",
    "prompt": "A 16:9 product poster featuring a futuristic glass API gateway, blue and violet volumetric lighting",
    "size": "2K",
    "output_format": "png",
    "response_format": "url",
    "watermark": false
  }'

Flash uses the same request shape:

{
  "model": "doubao-seedream-5-0-flash-260915",
  "prompt": "A 3:2 editorial photo of a Shanghai street corner after rain, neon reflections, cinematic natural light",
  "size": "1.5K",
  "output_format": "jpeg",
  "watermark": false
}

Single-Reference Image Generation

{
  "model": "doubao-seedream-5-0-pro-260628",
  "prompt": "Keep the pose and composition, but replace the silver garment with transparent glass",
  "image": "https://example.com/model.png",
  "size": "2K",
  "output_format": "png",
  "watermark": false
}

Multi-Reference Image Generation

{
  "model": "doubao-seedream-5-0-flash-260915",
  "prompt": "Keep the person from image 1 and replace their outfit with the clothes from image 2",
  "image": [
    "https://example.com/person.png",
    "https://example.com/clothes.png"
  ],
  "size": "2K",
  "output_format": "png",
  "watermark": false
}

Interactive Editing

Both Pro and Flash support interactive editing. Provide a marked-up reference or use normalized <point> and <bbox> coordinates in the prompt.

{
  "model": "doubao-seedream-5-0-pro-260628",
  "prompt": "Move the subject from image 1 <bbox>179 283 796 986</bbox> into image 2 <bbox>118 331 933 871</bbox>",
  "image": [
    "https://example.com/subject.png",
    "https://example.com/background.png"
  ],
  "size": "2K",
  "output_format": "png",
  "watermark": false
}

Layer Decomposition

Both Pro and Flash support layer decomposition. Set layer_decomposition: true to split the input image into a base image and up to 16 transparent PNG layers, which is useful for re-editing posters and covers. Use <bbox> coordinates (normalized to 0-1000) in the prompt to mark the text or subjects to extract.

{
  "model": "doubao-seedream-5-0-pro-260628",
  "prompt": "Split the image into precise layers. Text at <bbox>180 64 812 198</bbox> and <bbox>757 210 939 280</bbox>; the parrot at <bbox>347 305 642 997</bbox>.",
  "image": "https://ark-project.tos-cn-beijing.volces.com/doc_image/seedream_50_pro_layer_input.png",
  "layer_decomposition": true,
  "size": "2K",
  "output_format": "jpeg",
  "response_format": "url",
  "watermark": true
}

data[] is returned in layer order. The item with z_index: 0 is the base image; each extracted layer adds bounding_box.absolute (pixel [x1, y1, x2, y2]), bounding_box.normalized (0-1000), name, and description. Extracted layers are always png with transparency, and usage.generated_images counts the base image plus every layer.

{
  "url": "https://.../layer_1.png",
  "size": "2406x498",
  "output_format": "png",
  "z_index": 1,
  "bounding_box": {
    "absolute": [384, 120, 1653, 382],
    "normalized": [188, 59, 807, 186]
  },
  "name": "Seedream title text",
  "description": "The yellow Seedream title text without background elements"
}

Transparent Background

Set background: "transparent" to output an image with an alpha channel. This works only for image-to-image requests with exactly one input image that has an alpha channel (such as png). Output defaults to png; combining it with output_format: "jpeg" or a jpeg input returns an error.

{
  "model": "doubao-seedream-5-0-flash-260915",
  "prompt": "Keep the transparent background and turn the icon into frosted glass with soft highlights",
  "image": "https://example.com/icon-transparent.png",
  "background": "transparent",
  "size": "1K",
  "output_format": "png",
  "watermark": false
}

Base64 Response

Set response_format to b64_json; read the result from data[].b64_json.

{
  "model": "doubao-seedream-5-0-pro-260628",
  "prompt": "A minimalist blue mechanical whale icon on a white background",
  "size": "1K",
  "response_format": "b64_json",
  "output_format": "png",
  "watermark": false
}

Prompt Optimization

Pro supports standard and fast; Flash currently supports only standard.

{
  "model": "doubao-seedream-5-0-pro-260628",
  "prompt": "A futuristic city at night",
  "size": "2K",
  "optimize_prompt_options": {"mode": "fast"},
  "output_format": "png",
  "watermark": false
}

Images are returned in data[] through either url or b64_json. An item may contain an error; inspect every item instead of relying only on the HTTP status.