Qwen Image

Models

Availability and pricing differ by region. Check the model list for this endpoint before integrating — see the API catalog.

Both models serve text to image (T2I) and image to image / editing (I2I): no reference image means T2I, 1–3 reference images mean I2I. Images are returned synchronously — no polling.

The async task lane is not supported. This family is served only on the synchronous endpoints; a call to /v1/content_generation/tasks is refused.

Two inbound lanes

One model, two spellings, identical capability. Pick whichever matches the code you already have:

The two lanes spell size differently and neither accepts the other's spelling: the OpenAI-compatible lane uses the letter x (1024x1024), the Bailian native lane uses an asterisk (1024*1024). The wrong one returns 400 naming the separator it expects; it is not silently rewritten — mixing the two protocols should surface here rather than as a harder problem further down.

1. OpenAI-compatible lane

POST https://api.modelgo.ai/v1/images/generations

Text to image

curl https://api.modelgo.ai/v1/images/generations \
  -H "Authorization: Bearer $MODELGO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen-image-3.0-pro",
    "prompt": "A fluffy ginger cat on a sunlit windowsill, afternoon light, film grain",
    "size": "1024x1024",
    "n": 1
  }'

Image to image / editing

No multipart: the reference images go in the top-level image field.

curl https://api.modelgo.ai/v1/images/generations \
  -H "Authorization: Bearer $MODELGO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen-image-3.0-pro",
    "prompt": "Keep her facial features; change the outfit to a champagne silk shirt and a dark grey blazer",
    "image": ["https://example-cdn.com/portrait.png"],
    "size": "1024x1024",
    "prompt_extend": true
  }'

Response

{
  "created": 1788339600,
  "data": [{"url": "https://dashscope-result-sz.oss-cn-shenzhen.aliyuncs.com/xxx.png?Expires=xxx"}],
  "usage": {
    "output_width": 1024, "output_height": 1024,
    "input_image_count": 0, "input_image_type": "qima_input_1k",
    "output_image_count": 1, "output_image_type": "qima_output_1k"
  }
}

• Images are PNG, and the link is valid for 24 hours — download promptly and do not persist the URL.
• usage carries image counts and tiers, not token usage: this family bills purely per image (see Billing).
• response_format: "b64_json" does not error but is ignored; the response still carries an image URL.

Using the OpenAI SDK

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MODELGO_API_KEY"],
    base_url="https://api.modelgo.ai/v1",
    timeout=600.0,   # Generation is slow; use a generous timeout
)

resp = client.images.generate(
    model="qwen-image-3.0-pro",
    prompt="A fluffy ginger cat on a sunlit windowsill, film grain",
    size="1024x1024",
    n=1,
    extra_body={"prompt_extend": True},   # extensions go through extra_body
)
print(resp.data[0].url)

image / negative_prompt / seed / prompt_extend / prompt_extend_mode / enable_thinking / watermark are not OpenAI parameters: pass them via extra_body with the OpenAI SDK, or at the top level of the request body over plain HTTP.

2. Bailian native lane

POST https://api.modelgo.ai/api/v1/services/aigc/multimodal-generation/generation

Request and response are Bailian's own shapes in both directions — the gateway does not translate.

curl https://api.modelgo.ai/api/v1/services/aigc/multimodal-generation/generation \
  -H "Authorization: Bearer $MODELGO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen-image-3.0-pro",
    "input": {
      "messages": [
        {
          "role": "user",
          "content": [
            {"image": "https://example-cdn.com/portrait.png"},
            {"text": "Change the outfit to a champagne silk shirt"}
          ]
        }
      ]
    },
    "parameters": {"size": "1024*1024", "n": 1, "prompt_extend": true}
  }'

Response

{
  "output": {
    "choices": [
      {
        "finish_reason": "stop",
        "message": {
          "role": "assistant",
          "content": [{"image": "https://dashscope-result-sz.oss-cn-shenzhen.aliyuncs.com/xxx.png?Expires=xxx"}]
        }
      }
    ]
  },
  "usage": {
    "output_width": 1024, "output_height": 1024,
    "input_image_count": 1, "input_image_type": "qima_input_1k",
    "output_image_count": 1, "output_image_type": "qima_output_1k"
  },
  "request_id": "571ae02f-5c9d-436c-83c2-f221e6df0xxx"
}

Constraints of this lane

• input.messages carries exactly one object and role must be user (single-turn only).
• content carries exactly one {"text": ...}; omitting it or passing several returns 400.
• {"image": ...} 0–3 entries: 0 is text to image, 1–3 is image to image.
• Generation parameters live under parameters (not at the top level): size / n / seed / negative_prompt / prompt_extend / prompt_extend_mode / enable_thinking / watermark.
• With the official Bailian SDK (dashscope), point dashscope.base_http_api_url at https://api.modelgo.ai/api/v1 and use your ModelGo API key.

3. Parameters

Validation happens before the call

The following are refused at creation (HTTP 400) — nothing reaches the upstream and nothing is charged:

• n is not an integer in 1–6, or is written as a string or a decimal;
• size is far too small or large, or the aspect ratio is extreme; see the official documentation for the accepted range;
• size uses the other lane's separator (the error names whether x or * is expected);
• image is null, an empty array, or more than 3 entries;
• prompt_extend_mode: "agent" is combined with image (APE is text to image only);
• seed is outside 0 – 2147483647;
• on the native lane, messages is not exactly one, or content does not carry exactly one text.

Reference image requirements

• Formats: JPG / JPEG / PNG / BMP / TIFF / WEBP / GIF.
• Width and height between 384 and 2048 pixels recommended; at most 10 MB each.
• Public URLs (HTTP / HTTPS) or Base64 (data:{MIME_type};base64,{base64_data}).
• A public URL must be directly GET-able by the upstream: no auth, no redirect, no Referer / Cookie dependency. Some sites return 403 to non-browser clients, which fails the request.

4. Metering

This family is metered per image, not per token. The usage block carries image counts and tiers rather than token counts:

The tier is decided by the OUTPUT pixel area: an area of 2,250,000 (that is, 1500×1500) or less is 1k, anything larger is 2k. The input tier is judged by the output area too, not by the size of the reference image — that is the upstream's own rule.

Text to image supplies no reference image, so input_image_count is 0 and only the delivered images are metered. Requesting n images meters n, not one.

5. Full example (Python, official Bailian SDK)

import os
import dashscope
from dashscope import MultiModalConversation

dashscope.base_http_api_url = "https://api.modelgo.ai/api/v1"

response = MultiModalConversation.call(
    api_key=os.environ["MODELGO_API_KEY"],
    model="qwen-image-3.0-pro",
    messages=[{
        "role": "user",
        "content": [
            {"image": "https://example-cdn.com/portrait.png"},
            {"text": "Change the outfit to a champagne silk shirt; modern minimalist cafe background"},
        ],
    }],
    prompt_extend=True,
)

if response.status_code == 200:
    print(response.output.choices[0].message.content[0]["image"])
else:
    print(f"Error: {response.code} - {response.message}")