Image and Video Generation on Azure: gpt-image, Sora and Inpainting (AI-103)

The computer vision domain on AI-103 starts with generating visual content: images and videos from prompts and reference media, plus editing. Foundry offers image models such as gpt-image-1 and video models such as Sora, deployed like any other model.

Generate images from text

import base64

openai = project.get_openai_client()

result = openai.images.generate(
    model="image-main",            # your gpt-image deployment
    prompt="Flat illustration of a friendly robot sorting invoices on a desk, brand colours indigo and white",
    size="1536x1024",
    quality="high",
    n=1,
)
with open("robot.png", "wb") as f:
    f.write(base64.b64decode(result.data[0].b64_json))

Controls you can choose:

Better prompts are specific: subject, style, composition, lighting, colours and any text that must appear.

Generate from reference images

Pass reference images to keep a product, character or style consistent. The edit endpoint accepts one or more input images plus a prompt:

result = openai.images.edit(
    model="image-main",
    image=[open("product.png", "rb"), open("style-reference.png", "rb")],
    prompt="Place the product from the first image on a kitchen counter, in the style of the second image",
)

Edit images: inpainting and prompt-driven changes

Inpainting changes only part of an image. Supply a mask: a PNG the same size as the image, where the transparent area is what may change.

result = openai.images.edit(
    model="image-main",
    image=open("living-room.png", "rb"),
    mask=open("sofa-mask.png", "rb"),
    prompt="Replace the sofa with a green velvet sofa",
)

Edit workflows:

• Mask-based edits: replace or remove an object, or extend the background
• Prompt-driven edits: "make it night time", "change the text on the sign to OPEN", with no mask
• Iterative edits: feed the previous result back in for step-by-step refinement

Generate videos

Video generation is asynchronous, because rendering takes time:

1. Create a job with a prompt, plus optional reference image or video, duration, resolution / aspect ratio and number of variants
2. Poll the job status until it succeeds
3. Download the video (and often a thumbnail)

job = openai.videos.create(model="video-main", prompt="Drone shot over a foggy pine forest at sunrise", seconds="8", size="1280x720")
while job.status not in ("completed", "failed"):
    time.sleep(10)
    job = openai.videos.retrieve(job.id)
content = openai.videos.download_content(job.id)

Endpoint names and parameters differ between video model versions and API versions; check the current docs. The create → poll → download pattern is what matters.

Edit generated videos

• Remix: create a new version of a generated video with a changed prompt ("same scene, but in snow"), keeping structure and motion
• Image-to-video: animate a still image used as the first frame
• Extend or trim, and combine clips in your own pipeline (ffmpeg or a media service)

Choosing controls

Safety and provenance

Image and video generation runs through content filters on the prompt and the output, and refuses disallowed requests such as sexual content, graphic violence, or realistic images of real public figures where policy forbids it. Generated media carries content credentials (C2PA) that mark it as AI-generated. Add your own checks for brand rules before publishing (covered in the next lesson).

Key takeaways

• Images: images.generate(model, prompt, size, quality, n); edits with reference images and masks via images.edit.
• Inpainting: the mask's transparent area is what changes.
• Video: asynchronous job → poll → download; remix and image-to-video for edits.
• Choose size, quality, count, duration and references to fit the use; generation is filtered and marked with C2PA credentials.

Next lesson: multimodal understanding and visual safety.