MAI Image family

Microsoft’s image generation and editing models, three in total, called through the OpenAI-compatible API in much the same way as GPT Image: text to image uses /v1/images/generations, and editing uses /v1/images/edits.

Choosing a model

All three take the same parameters; in our tests a 1024×1024 image takes about 12 to 25 seconds. For live prices, see the model page .

Text to image

import base64
from openai import OpenAI
 
client = OpenAI(api_key="YOUR_OFOX_API_KEY", base_url="https://api.ofox.ai/v1", timeout=600)
 
resp = client.images.generate(
    model="microsoft/mai-image-2.5-flash",
    prompt="A red apple on a white table",
    size="1024x1024",
)
 
with open("output.png", "wb") as f:
    f.write(base64.b64decode(resp.data[0].b64_json))

curl https://api.ofox.ai/v1/images/generations \
  --max-time 600 \
  -H "Authorization: Bearer $OFOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "microsoft/mai-image-2.5-flash",
    "prompt": "A red apple on a white table",
    "size": "1024x1024"
  }' | jq -r '.data[0].b64_json' | base64 --decode > output.png

Size (size)

• The output keeps the requested aspect ratio, with total pixels adjusted to about 1 million; a size that is too small produces 768×768.
• For higher resolutions, use Nano Banana Pro (up to 4K) or GPT Image.

Image editing

curl https://api.ofox.ai/v1/images/edits \
  --max-time 600 \
  -H "Authorization: Bearer $OFOX_API_KEY" \
  -F model=microsoft/mai-image-2.5 \
  -F prompt="Make the apple purple, keep everything else the same" \
  -F image=@apple.jpg \
  | jq -r '.data[0].b64_json' | base64 --decode > apple_purple.png

Reference images count as input tokens; see usage.input_tokens_details.image_tokens in the response. We measured 576 tokens for one 768×768 reference image.

Common errors

For multiple images, send several requests concurrently, one image per request.

FAQ