编辑图像

请求体与生成图像完全相同，唯一区别是必须提供至少一张参考图，通过 image 字段传入。

用 multipart/form-data 上传源图文件，或用 JSON 把 image 设为 URL 或 data: URI。

本接口没有独立的蒙版字段。需要只改局部时，改用坐标编辑：在 prompt 里写 <bbox> 或 [x1, y1, x2, y2] 指定区域。

The body is identical to Generate image; the one difference is that at least one reference image must be supplied through image.

Upload the source file with multipart/form-data, or set image to a URL or data: URI in JSON.

This endpoint has no dedicated mask field. To change only part of the image, use a coordinate edit instead: put <bbox> or [x1, y1, x2, y2] in prompt.

请求体可以走 multipart/form-data，用文件而非 URL 提交底图。所有标量字段都是字符串。

curl --request POST \
  --url 'https://ark.oairoute.com/v1/images/edits' \
  --header 'Authorization: Bearer YOUR_ARK_KEY' \
  --form 'model=doubao-seedream-5-0-pro-260628' \
  --form '[email protected]' \
  --form 'prompt=Turn this photo into a night scene' \
  --form 'size=2K'

The body can use multipart/form-data to send the base image as a file rather than a URL. Every scalar field is a string.

curl --request POST \
  --url 'https://ark.oairoute.com/v1/images/edits' \
  --header 'Authorization: Bearer YOUR_ARK_KEY' \
  --form 'model=doubao-seedream-5-0-pro-260628' \
  --form '[email protected]' \
  --form 'prompt=Turn this photo into a night scene' \
  --form 'size=2K'

本接口没有 mask。要只改画面中的一块区域，把区域写进提示词：

• 框选：把框 [225, 442, 796, 1414] 内的区域替换成一把红色纸伞
• 或在图片上先画好标记再上传，用自然语言描述要改的地方

坐标是该图的像素坐标，原点在左上角。模型只处理描述到的区域，其余保持不变。

This endpoint has no mask. To change only one region, describe the region in the prompt:

• Box selection: Replace the area inside [225, 442, 796, 1414] with a red paper umbrella
• Or draw marks on the image first and upload it, then describe the change in words

Coordinates are that image's pixels with the origin at the top-left. The model touches only the described region and leaves the rest alone.

提示词描述目标状态

与其写「把红衣服改成蓝衣服」，不如写「人物穿蓝色外套，其余不变」。

保留构图

希望整体风格不动、只微调细节时，在提示词里强调「保持构图、光线与人物不变」。

多轮编辑会累积失真

每次编辑都重新编码，反复五次以上画质会明显下降。需要多步修改时尽量一次说清，或从原图重新开始。

Describe the target state

Rather than "change the red shirt to blue", write "the person wears a blue coat; everything else unchanged".

Preserve the composition

When the overall look should stay and only details change, stress "keep the composition, lighting, and subject unchanged".

Repeated edits accumulate loss

Each edit re-encodes the image; beyond about five rounds quality drops noticeably. Batch changes into one prompt or restart from the original.

同步返回，结构与生成图像一致：

{
  "data": [
    {
      "url": "https://ark-content-generation-v2-cn-beijing.tos-cn-beijing.volces.com/seedream/edit-xxx.jpeg",
      "size": "2048x2048",
      "output_format": "jpeg"
    }
  ],
  "created": 1757469067,
  "usage": {
    "generated_images": 1,
    "input_images": 1,
    "output_tokens": 15908,
    "total_tokens": 15908
  }
}

• data[].url (string) — 图片地址，response_format 为 url 时返回，24 小时有效。
• data[].b64_json (string) — response_format 为 b64_json 时返回。
• data[].size (string) — 实际像素尺寸。
• data[].output_format (string) — 文件格式。
• usage.input_images (integer) — 实际接收的参考图张数。

The response is synchronous and shaped like Generate image:

{
  "data": [
    {
      "url": "https://ark-content-generation-v2-cn-beijing.tos-cn-beijing.volces.com/seedream/edit-xxx.jpeg",
      "size": "2048x2048",
      "output_format": "jpeg"
    }
  ],
  "created": 1757469067,
  "usage": {
    "generated_images": 1,
    "input_images": 1,
    "output_tokens": 15908,
    "total_tokens": 15908
  }
}

• data[].url (string) — the image link, returned when response_format is url, valid 24 hours.
• data[].b64_json (string) — returned when response_format is b64_json.
• data[].size (string) — actual pixel size.
• data[].output_format (string) — the file format.
• usage.input_images (integer) — how many reference images were actually received.

常见 400 原因：没有传 image；image 不是可访问的地址或 Base64 格式标识未小写；n 大于 1；参考图超过 10 张。

更多状态码与重试建议见错误码与限流，常见报错的排查步骤见常见问题。

Common 400 causes: no image; an unreachable URL or a Base64 format not in lowercase; n above 1; more than 10 reference images.

For more status codes and retry guidance see Error codes and rate limits; for troubleshooting steps see the FAQ.

以下图像模型均支持参考图编辑，清单只列各厂商最新一代，实际可用范围以你的账号分组为准：

Every image model below supports reference-image editing; the list shows only each vendor's latest generation, and what your account group actually allows takes precedence:

整体重绘

只传原图与提示词，模型会在保留主体构图的前提下重画整张图。适合换风格、换材质这类改动。

返回 Base64

需要把结果直接存进数据库或内网服务时，用 b64_json 避免二次下载。

Full redraw

Send the source image and a prompt; the model repaints the whole frame while keeping the subject composition. Good for style or material changes.

Base64 output

When results go straight into a database or an internal service, use b64_json to avoid a second download.