Image Generation Overview

QWave API supports multiple image generation models, all called through the unified async task endpoint /v1/tasks. Image generation typically completes in 1-30 seconds. After submission, you receive a task_id and retrieve results via polling or webhook callback.

Endpoint

POST https://api.qingbo.ai/v1/tasks

For the async flow, see Task System — Submit → Status → Result + optional webhook.

Supported Models

Gemini Image Series

The three Nano Banana official tiers and Nano Banana 2.1 billed from actual token usage, five economy tiers at a fixed price per image, up to 14 reference images

GPT Image Series

GPT Image 1 / 1.5 / 2 / 2.5 official lines billed from actual token usage, GPT Image 2 / 2.5 reverse lines at a fixed price per image

Seedream Series

ByteDance Seedream 4.0 / 4.5 / 5.0 Lite / 5.0 Pro / 5.0 Flash, text-to-image and reference-image editing

Qwen Image Series

Alibaba Qwen Image 2.0 / 2.0 Pro / 3.0 / 3.0 Pro, strong text rendering

Wan Image Series

Alibaba Wan 2.7 Image / Image Pro: text-to-image, reference editing, bbox selection and color palettes

Z-Image

Z-Image-Turbo, lightweight and fast, bilingual CN/EN

FLUX Series

Black Forest Labs FLUX.2 Pro / Max / Flex and Kontext Pro / Max, text-to-image and reference-image editing

FLUX 3 Image

Black Forest Labs FLUX 3 Image: text-to-image, single-image editing and up to 10 reference images, up to 4K

Grok Imagine Series

xAI Grok Imagine 1.5 reverse tiers, 2.0 official and reverse tiers, and the quality tier: text-to-image and editing

Midjourney

imagine returns four images per call, plus 12 chained actions: upscale, variation, zoom, pan, inpaint and more

MAI-Image-2.6

Microsoft MAI-Image-2.6 and Flash: text-to-image, single-image editing and composition from up to 5 references, billed from actual token usage

Two request paths

Images have two entries. Billing is identical on both; they differ in parameter validation and how results come back:

	`POST /v1/tasks`	`POST /v1/images/generations` · `/v1/images/edits`
Result	Async — you get a `task_id` and poll (or wait inline with `Prefer: wait`)	Sync — the image comes back in the response
Request shape	This platform’s unified fields: `action` / `aspect_ratio` / `resolution` / `image_urls`	OpenAI’s native shape: `size` / `quality` / `n` / `response_format`
Validation	Full — action, aspect ratio, resolution, quality and reference-image count are all checked; a non-conforming request returns `400` and is not billed	The compatibility layer does not run these checks, so invalid parameters are not caught up front
Use it for	New integrations, and anything needing masked inpainting, quality tiers, resolution tiers or multi-image references	Existing OpenAI SDK projects — point `base_url` at us and go

To migrate with the OpenAI SDK, set base_url to https://api.qingbo.ai/v1 and change model to one of our model IDs. Model-specific capabilities (the mask_url mask, the 1k / 2k / 4k resolution tiers, reference-image limits and so on) follow the task API documentation — the parameters on our model pages describe the task API.

Common Parameters (shared across models)

Actual support varies per model — see each vendor doc for details.

string

required

Model ID (group_name); pick one from the vendor docs above

string

default:"generate"

generate — text to image (default)
edit — image editing; pair with image_urls

The mode is decided by what you send, not by the action: prompt alone is text-to-image; one image_urls entry is image-to-image; several are multi-image reference; add mask_url for inpainting.

string

required

Image description text; supports both Chinese and English

integer

default:"1"

Number of images to generate (some vendors cap per-call count)

integer

default:"-1"

Random seed; -1 means random. A fixed value reproduces similar results.

string

Aspect ratio, e.g. 16:9 / 9:16 / 1:1. Some vendors support auto (smart selection).

string

Output resolution, e.g. 1K / 2K / 4K. Supported range varies per vendor.

string[]

Reference image URL array (for image-to-image / multi-image reference)

string

Webhook callback URL, invoked when the task reaches a terminal state. See Task System.

Submit Response Example

{
  "task_id": "task-wave1775285160b950328499",
  "model": "seedream-5.0-lite",
  "action": "generate",
  "status": "queued",
  "created_at": 1775285160040,
  "progress": 0
}

After receiving the task_id, call GET /v1/tasks/{task_id} to poll until status = completed, then read the image URLs.

Mode Quick Reference

Mode	Trigger Fields	Typical Vendors
Text-to-image (T2I)	`prompt` only	All
Image-to-image / multi-image reference	`+ image_urls` (single or multiple)	Seedream / Gemini / GPT Image / Wan / Qwen / FLUX / Grok
Image editing (inpainting / lighting changes)	`+ image_urls + action: edit`	Seedream / Gemini / GPT Image / Wan / Qwen / FLUX / Grok
Inpainting (with mask)	`+ image_urls + mask_url + action: edit`	The GPT Image official models; not supported on the GPT Image reverse lines or the Nano Banana economy routes
Sequential image group (thematic set)	`enable_sequential: true`	Wan Image
Interactive editing (bbox selection)	`+ image_urls + bbox_list`	Wan Image

Not every model supports every mode — check each vendor doc for the actual action list and field range. The backend validates request fields against the vendor’s declared capabilities and rejects out-of-range requests.

Field Naming Conventions

Media references are always plural — always image_urls, even for a single image use a one-element array ["one.jpg"]
Aspect ratio is unified as aspect_ratio — vendor internals using size / ratio are implementation details you don’t need to track
Resolution is unified as resolution — vendor internals using quality are implementation details
Mask field mask_url (used for GPT Image official inpainting)

Task System — task state machine / polling cadence / webhook
Request & Response — common error codes / headers / rate limits
Authentication — API key application and usage
Models — full model lookup endpoint

Text Generation