Nova Canvas outpainting DEFAULT vs PRECISE, and a Sume ratio edit

Outpainting means making a picture wider or taller than the photo you have. Amazon Nova Canvas has a dedicated OUTPAINTING task for it, with a mode switch. Sume has no outpaint task, so the question is what to send instead and how to check the seam.

The two Nova Canvas modes

Per Amazon's docs (read 2026-10-06), outPaintingMode is DEFAULT or PRECISE. DEFAULT can leave a visible halo around the original image. PRECISE adheres strictly to the mask. Outpainting also reverses the inpainting mask convention, so check the user guide before you build the mask.

What Sume does

On Sume you send the photo in input_references and set an aspect_ratio that differs from the source. The Image API docs say an Ideogram 4.5 edit with no aspect_ratio keeps the source shape, which means that setting one is how you change it. The docs do not promise that the original pixels stay unchanged in the middle, so treat the result as a new render and test it.

For models with aspect_ratio in their catalog row you can read the allowed values from GET /v1/images/models/{id}/endpoints.

Test the seam instead of trusting a mode

Crop the centre of the output at the source's aspect and compare it with the source. A mean absolute difference near zero means the middle was kept. A larger number means the model repainted it, and you should not use that route for packshots where the product must not change.

import io, os, requests
import numpy as np
from PIL import Image

SRC = "https://example.com/shoe.jpg"
r = requests.post("https://api.sume.com/v1/images",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    json={"model": "ideogram/ideogram-v4.5", "aspect_ratio": "16:9",
          "prompt": "Extend the scene to the left and right. Do not change the shoe.",
          "input_references": [{"type": "image_url", "image_url": {"url": SRC}}]},
    timeout=90)
r.raise_for_status()
src = Image.open(io.BytesIO(requests.get(SRC).content)).convert("RGB")
out = Image.open(io.BytesIO(requests.get(r.json()["data"][0]["url"]).content)).convert("RGB")
w, h = out.size
sw = round(h * src.width / src.height)
x0 = (w - sw) // 2
mid = out.crop((x0, 0, x0 + sw, h)).resize(src.size)
diff = np.abs(np.asarray(mid, float) - np.asarray(src, float)).mean()
print("mean abs diff:", round(diff, 2))

Run it on five of your own photos before you pick a route. The result depends on the picture, not on the mode name.