Skip to main content

Image Generation and Editing

Level: Beginner · Category: API reference

Tags: Text-to-image · Image editing

POST https://openai.rc.asu.edu/v1/images/generations makes images from a text prompt, and POST /v1/images/edits changes an image you upload, in the OpenAI Images API shape. Keys and setup are on LLM API Access.

ModelNotes
flux-2FLUX.2 [dev], 32B. Text-to-image and edits, including edits guided by several reference images.
qwen-image-21Qwen-Image 2.1. Text-to-image and edits; licensed for non-commercial research use only.

Check each model's license (in its description in /v1/models or Voyager) before using images outside research.

Generate an image​

import base64
from openai import OpenAI

client = OpenAI()  # reads OPENAI_API_KEY and OPENAI_BASE_URL
result = client.images.generate(
    model="flux-2",
    prompt="a saguaro cactus at sunrise, photorealistic",
    size="1024x1024",
)
with open("saguaro.png", "wb") as f:
    f.write(base64.b64decode(result.data[0].b64_json))
  • The image comes back base64-encoded in data[0].b64_json (PNG). The response also has the seed that was used.
  • size is WIDTHxHEIGHT, for example 512x512 or 1024x1024. Larger images take longer: in testing, a 512x512 image took about 30 seconds and a 1024x1024 image about a minute.
  • Set the client timeout to several minutes; when many people generate images at once, requests wait for a free accelerator (see Limits and errors).

Edit an image​

Upload the image as a file with a prompt that describes the change:

import base64
from openai import OpenAI

client = OpenAI()
with open("saguaro.png", "rb") as image:
    result = client.images.edit(
        model="flux-2",
        image=image,
        prompt="add a full moon in the sky",
        size="1024x1024",
    )
with open("saguaro-moon.png", "wb") as f:
    f.write(base64.b64decode(result.data[0].b64_json))

Edits are multipart uploads and aren't retried automatically, so retry in your own code if one fails with a 502 or 503.

Images inside a chat​

Some chat models can also draw during a conversation: the gateway gives them a generate_image tool (the image_generation add-on, see Gateway add-ons), and the image is appended to the reply as a Markdown data URL. For scripts, call /v1/images/generations directly.