**Cost**: configured per generated image in `pricing.yaml` (multiplied by `n`).

Text-to-image generation. Produce one or more images from a natural-language prompt, optionally steered by reference images for style or composition. Synchronous; routes through whichever model you pick.

## What this API is for

The canonical "generate image from text" endpoint. Use it for hero art, marketing assets, agent-driven illustration, batched A/B variants. Reference images give you style/subject control without fine-tuning.

## Request

| Field | Required | Notes |
|---|---|---|
| `prompt` | yes | Text description of the image to generate. The most important field — be specific about subject, composition, lighting, and style. |
| `n` | no | Number of images to generate. Default `1`. **Credits are charged per generated image**, not per call. |
| `size` | no | Output dimensions as `WIDTHxHEIGHT`, e.g. `1024x1024`. Defaults vary by model. Common: `1024x1024`, `1536x1024`, `1024x1536`. |
| `model` | no | Model ID. Default `nano_banana_2`. Available: `nano_banana_2`, `nano_banana_pro`, `gpt_image_2_5`, `gpt_image`, `qwen_image_3_pro`, `hy_image_3_5`, `seedream_5_pro`, `seedream_5_flash`, `seedream_5`, `seedream_4_5`, `recraft_v4`, `flux_3`, `flux_2_pro`, `flux_2_max`, `ideogram_v4`. Each has its own aesthetic / strength — see the table below. |
| `quality` | no | For `gpt_image`, use `low`, `medium`, `high`, or `auto`. For `gpt_image_2_5` (including Flare and Sunburst), `xhigh` and `max` are also supported. Omitted quality or `auto` uses the medium billing bucket. For non-GPT models, this field may carry the resolution tier (for example `1K` or `2K`). |
| `background` | no | Background mode (model-dependent). For `gpt_image`: `transparent` / `opaque` / `auto`. |
| `reference_images` | no | Array of public image URLs to steer the generation. Most models accept up to a handful — semantics vary (style ref vs. subject ref). `flux_3` accepts up to 10 reference images. |

### Model cheat-sheet

| Model | Strength |
|---|---|
| `nano_banana_2` (default) | Fast, general-purpose, good prompt adherence. |
| `nano_banana_pro` | Higher fidelity at higher latency. |
| `gpt_image_2_5` | GPT Image 2.5 with Flare and Sunburst variants; supports low, medium, high, xhigh, max, and auto quality. |
| `gpt_image` | Best for text inside images and structured layouts. Supports `quality` and `background`. |
| `qwen_image_3_pro` | Detailed layouts, multilingual text rendering, dense information-rich images, and 512/1K/2K outputs. No 4K. `image_generate` accepts up to 3 references; `image_to_image` accepts up to 2. |
| `hy_image_3_5` | Photorealistic generation and reference-guided editing; 1K/2K outputs and up to 5 input images. |
| `seedream_5_pro` | Precision generation and editing with advanced visual reasoning; 1K/2K outputs and up to 10 references. |
| `seedream_5_flash` | Fast, low-cost generation and editing; 1K/2K outputs at one price and up to 10 references. |
| `seedream_5` / `seedream_4_5` | Cinematic, photographic. |
| `flux_3` | Native 1K/2K/4K generation, precise prompt following, and up to 10 references. |
| `flux_2_pro` / `flux_2_max` | High-quality, very precise prompt following. |
| `recraft_v4` | Vector-friendly art / illustration. |
| `ideogram_v4` | Latest Ideogram typography, posters, logos. |

## Response

Shared `ImageUrlListSuccess` — `data.image_urls` is an array of CDN URLs, length `n`.

## Billing

**Credits charged per generated image**, not per call. For `n = 4`, the configured per-image model price is multiplied by 4. Check `GET /v2/open/limits` for the live per-operation price table and `GET /v2/open/usage` for charged calls. Uses the unified Codia Open API balance.

## Errors

- `400` — missing/empty `prompt`, invalid `size` format, unknown `model`, too many `reference_images`.
- `401` / `403` — auth.
- `402` — credits exhausted.
- `429` — rate limit.
- `500` — internal; safe to retry.

> **Sync vs. async**
> This synchronous endpoint returns the result inline. When `n > 1` the call generates images concurrently and may still approach HTTP timeouts; prefer `POST /v2/open/tasks` with `operation=image_generate` for batch jobs (and for webhook delivery / retry semantics). Inputs and per-image credit cost are identical.

可选，由用户自定义的幂等键（最多 255 个字符）。设置后，成功（2xx）的响应会缓存 7 天。使用相同的键 + 相同的请求体重试，会返回缓存的响应（不额外扣除积分）；使用相同的键但不同的请求体重试，则返回 409。用于确保 POST 请求在网络故障时可以安全重试。

Idempotency-Key

prompt

要生成的图片数量。按生成的图片张数扣除积分。

输出图片尺寸，格式为 `WIDTHxHEIGHT`。常用尺寸：`1024x1024`、`1536x1024`、`1024x1536`。FLUX 2 Pro/Max 也接受 1K/2K 尺寸预设。超大请求（包括 4K）会被缩减到最多约 4MP（方形图片为 2048x2048）；不支持原生 4K 输出。计费按调整后的输出面积计算：总像素不超过 1,048,576 的按 1K 价格计费，更大的输出按 2K 价格计费。显式指定的 size 优先于 quality 中的分辨率提示。

size

生成图片的模型。默认为 `nano_banana_2`。支持的模型与 Codia Studio 的文生图一致。积分在 `pricing.yaml` 中配置，并通过 `GET /v2/open/limits` 返回。

| 模型 | 适用场景 |
| --- | --- |
| `nano_banana_2` | 快速的日常生成（Gemini 3.1 Flash）。 |
| `nano_banana_pro` | 具备高级推理能力的影棚级 4K 视觉（Gemini 3 Pro）。 |
| `gpt_image_2_5` | GPT Image 2.5，含 Flare 和 Sunburst 变体；支持 low、medium、high、xhigh、max 和 auto 质量。 |
| `gpt_image` | 图内文字清晰、参考图还原度高、尺寸灵活。支持 `quality`（`low`/`medium`/`high`/`auto`）和 `background`（`transparent`/`opaque`/`auto`）。 |
| `qwen_image_3_pro` | 精细排版、多语言文字渲染、信息密集的图片，支持 512/1K/2K 输出，不支持 4K。`image_generate` 最多接受 3 张参考图；`image_to_image` 最多接受 2 张。 |
| `hy_image_3_5` | 1K/2K 的照片级写实生成和参考图引导编辑，不支持 4K。每次请求最多接受 5 张输入图片。 |
| `seedream_5_pro` | 具备高级视觉推理能力的精准生成；支持 1K/2K 输出，最多 10 张参考图。 |
| `seedream_5_flash` | 快速、低成本的生成；1K/2K 输出同价，最多 10 张参考图。 |
| `seedream_5` | 细节丰富的照片级写实图像。尺寸以 1920 为基准，例如 `1920x1080`。 |
| `seedream_4_5` | 稳定可靠的写实效果，非常适合人像。尺寸以 1920 为基准。 |
| `recraft_v4` | 插画、图标和矢量风格作品，风格控制精准。 |
| `flux_3` | 原生 1K/2K/4K 生成，提示词遵循精准，最多 10 张参考图。 |
| `flux_2_pro` | Black Forest Labs 出品的高质量 FLUX 生成。 |
| `flux_2_max` | FLUX 最高质量档位，适合追求顶级效果。 |
| `ideogram_v4` | 最新的 Ideogram，擅长文字排版和创意构图。 |

model

对于 `gpt_image`，可使用 `low`、`medium`、`high` 或 `auto`。对于 `gpt_image_2_5`（包括 Flare 和 Sunburst），还支持 `xhigh` 和 `max`。省略 quality 或设为 `auto` 时，按 medium 计费档位计费。对于非 GPT 模型，此字段可能用于传递分辨率档位（例如 `1K` 或 `2K`）。

quality

背景模式（取决于模型）。对于 GPT Image 2 和 2.5：`transparent`、`opaque` 或 `auto`。

background

用于风格引导的参考图片 URL（取决于模型）

reference_images

code

message

image_urls

data

请求体格式错误或缺少必填参数

Idempotency-Key 已被用于不同的请求体。

请求频率超限。请等待 `Retry-After` 响应头中指定的秒数后再重试。

根据文本提示词生成图片，可选用参考图引导，支持多种模型。

图片生成

输入

输出

相关 API