生成图像

请求体与 POST /doubao/api/v3/images/generations 相同，区别只在路径与 response_format 的可用取值。

客户端已按 OpenAI Images API 编写时使用本路径。相比原生路径多两点能力：

• 支持 response_format: b64_json，网关会下载生成的图片并内联返回。
• 支持 multipart/form-data，参考图可以上传文件而不必用 URL。

n 必须为 1。需要多张请用 sequential_image_generation，不要再靠 n。

文生图、图生图还是多图融合，只由 image 字段决定：

The request body is identical to POST /doubao/api/v3/images/generations; only the path and the accepted response_format values differ.

Use this path when your client already speaks the OpenAI Images API. It adds two things over the native path:

• response_format: b64_json is supported; the gateway downloads the images and returns them inline.
• multipart/form-data is supported, so reference images can be uploaded as files instead of URLs.

n must be 1. Use sequential_image_generation for multiple images rather than n.

Text-to-image, single-image edit, or multi-image fusion is decided solely by image:

档位： 1K、1.5K、2K，默认 2K。

精确像素： 宽高各自须落在 1024 到 4096 之间，宽高比须在 1:16 到 16:1 之间。

限制的是单边与比例，不是乘积：1024x1024 合法，512x512 会返回 400。

Tiers: 1K, 1.5K, 2K; the default is 2K.

Exact pixels: width and height must each fall between 1024 and 4096, with an aspect ratio between 1:16 and 16:1.

The per-side bounds and ratio are what's limited, not the product: 1024x1024 is valid, while 512x512 returns 400.

同步返回，data 中每项对应一张生成的图片：

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

• data[].url (string) — 有效期 24 小时的图片链接，response_format 为 url 时返回。
• data[].b64_json (string) — response_format 为 b64_json 时返回的 Base64 内容。
• data[].size (string) — 该张图片的实际像素尺寸。
• data[].output_format (string) — 文件格式。图层固定为 png，底图跟随请求。
• data[].z_index / name / description / bounding_box — 仅图层拆分时返回。
• data[].error (object) — 单图失败信息，仅组图场景中某一张失败时出现。
• created (integer) — 完成时间，Unix 秒。
• usage (object) — 用量信息，含 generated_images、input_images、output_tokens、total_tokens。

The response is synchronous; each entry in data is one generated image:

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

• data[].url (string) — a link valid for 24 hours, returned when response_format is url.
• data[].b64_json (string) — inline Base64, returned when response_format is b64_json.
• data[].size (string) — the actual pixel size of that image.
• data[].output_format (string) — the file format. Layers are always png; the base image follows the request.
• data[].z_index / name / description / bounding_box — layer decomposition only.
• data[].error (object) — per-image failure, present only when one image in a group fails.
• created (integer) — completion time in Unix seconds.
• usage (object) — generated_images, input_images, output_tokens, total_tokens.

常见 400 原因：model 写了缩写；size 超出 1024x1024 到 4096x4096 或比例越界；n 大于 1；参考图超过 10 张；图层拆分未传图或传了多张；透明背景用于文生图。

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: an abbreviated model; size outside 1024x1024 to 4096x4096 or a bad ratio; n above 1; more than 10 reference images; layer decomposition with no image or several; a transparent background on text-to-image.

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

以下图像模型均可通过本接口提交任务，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：

两个型号的参数完全相同，差别只在速度与成本：

Every image model below can be submitted through this endpoint; the list shows only each vendor's latest generation, and what your account group actually allows takes precedence:

The two variants take identical parameters and differ only in speed and cost:

总价 = 输出图单价 + 参考图加价 × max(0, 参考图张数 − 1)。

输出图按实际像素面积分档，第 1 张参考图免费，第 2 张起每张加收固定单价。任务失败自动全额退款。图层拆分时每个输出图层单独计费。

具体价格见模型价格与分组接口。

Total = output-image price + reference surcharge × max(0, reference count − 1).

Output images are tiered by actual pixel area. The first reference image is free and each one after it adds a flat surcharge. Failed tasks are refunded in full. Layer decomposition bills every output layer separately.

See Model pricing and group API for actual prices.

内联 Base64 输出

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。

组图

sequential_image_generation: "auto" 一次最多返回 4 张同主题图片。某一张失败时该数组项只带 error，其余照常返回，usage.generated_images 只计成功的张数。

图层还原

拆分得到的图层按 z_index 从小到大叠放。绝对坐标换算为 x=left、y=top、w=right−left、h=bottom−top；任意 W×H 画布用归一化值除以 1000 再乘画布尺寸。

Inline Base64 output

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

Image groups

sequential_image_generation: "auto" returns up to 4 images on one theme. If one fails, that array entry carries only error while the rest still return, and usage.generated_images counts only the successes.

Layer reconstruction

Stack the returned layers by ascending z_index. Convert absolute boxes with x=left, y=top, w=right−left, h=bottom−top; on an arbitrary W×H canvas, divide the normalized values by 1000 and multiply by the canvas size.