AI image aspect ratio changer: re-render at 9:16 or 16:9

To change an image's aspect ratio with AI without cropping the subject out, either extend the canvas and let a model fill only the new area (often called outpainting), or have an image model redraw the whole picture at the new ratio from your original. Outpainting leaves the original area in place; a redraw can shift details anywhere in the frame. If you can lose the edges, a plain crop in an image editor changes the ratio with no AI at all.

Sume has no expand or uncrop tool today, so on Sume this is the redraw: a POST /v1/images call with the original as a reference and the new aspect_ratio. The details below come from the Image API docs and the ratio lists GET /v1/images/models serves, read on 2026-09-27.

How do I change an image's aspect ratio with the Image API?

Send the original in input_references, set aspect_ratio to the new shape rather than "auto" (which matches the reference), and say in the prompt what should fill the new space.

• The ratio must be on the model's list; a model only accepts the values its catalog lists.
• The original must be at a public HTTPS URL. Localhost, private-network, and non-HTTPS URLs are rejected before submission.
• Ask for a few candidates with n, up to 4 per call on these models except Seedream 4.5, which today returns one image per call when you send a reference.

curl -X POST "https://api.sume.com/v1/images" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/nano-banana-2",
    "prompt": "The same scene as the reference image, reframed as a tall picture: keep the subject, style, and colors, and continue the sky above and the street below.",
    "aspect_ratio": "9:16",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/square-original.jpg" } }
    ]
  }'

Which ratios can I change an image to?

Each model publishes its own list, and today the lists differ at the extremes: Nano Banana 2 adds the banner shapes 4:1, 1:4, 8:1, and 1:8, while the ChatGPT Image models also take custom pixel sizes. For these models that edit from a reference:

Is this the same as expanding or uncropping an image?

No. An image model generates the whole picture again from the reference and your prompt rather than extending the original pixels, so faces, lettering, and small details can change. Compare every result with the original before you use it, and describe what belongs in the new space rather than leaving it to the model.

Generating new images at a chosen ratio or a custom pixel size from the start is covered in AI image generation API aspect ratios.

What are the limits?

• A text-to-image-only model, whose input_references descriptor is {"min": 0, "max": 0}, rejects the original.
• Each completed image is billed at the model's catalog rate, and a failed generation is not billed.
• Results are Sume-hosted, signed URLs. Download the versions you keep.
• A ratio is not a pixel size. To reach exact pixels afterwards, see upscale an image to an exact pixel size.
• This is for stills. To change a video's shape, see convert a landscape video to vertical.