Check an AI pattern tiles seamlessly with ImageChops.offset

Wrap the image by half its width and height with ImageChops.offset(img, w // 2, h // 2), which moves the old edges to the center cross, then compare pixels across that cross with their neighbors. A tile that repeats cleanly has no visible step there. A tile that does not has a hard line, and the score below turns that line into a number.

Why prompts do not guarantee a tile

Asking for a seamless pattern makes an image model draw a pattern, but not one whose left edge matches its right edge. Some models come close, many do not. Sume's image models do not have a tile parameter in the catalog, so you check the result yourself.

The offset trick is the standard check in texture work: put the seam where you can see it. If you can see it, a visitor will, once the tile repeats across a page.

Score the seam

After the offset, the old edges sit on the middle row and middle column. The code takes a thin strip across each and compares it with a strip just beside it using ImageChops.difference, then reads the mean with ImageStat. The comparison strip is 4 pixels from the seam.

import os, io, requests
from PIL import Image
H = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}
def gen(**body):
    r = requests.post("https://api.sume.com/v1/images", json=body, headers=H, timeout=60)
    r.raise_for_status()
    if r.status_code == 202:
        raise SystemExit("queued, read /v1/jobs/{id}/result: " + r.text)
    return r.json()
def fetch(u):
    return Image.open(io.BytesIO(requests.get(u, timeout=60).content))
from PIL import ImageChops, ImageStat
out = gen(model="black-forest-labs/flux.2-pro", aspect_ratio="1:1",
          prompt="Seamless repeating pattern of small blue paper boats on cream, flat design")
img = fetch(out["data"][0]["url"]).convert("L")
w, h = img.size
sh = ImageChops.offset(img, w // 2, h // 2)
def strip_diff(box_a, box_b):
    d = ImageChops.difference(sh.crop(box_a), sh.crop(box_b))
    return ImageStat.Stat(d).mean[0]
vert = strip_diff((w // 2 - 1, 0, w // 2 + 1, h), (w // 2 + 3, 0, w // 2 + 5, h))
horz = strip_diff((0, h // 2 - 1, w, h // 2 + 1), (0, h // 2 + 3, w, h // 2 + 5))
base = ImageStat.Stat(ImageChops.difference(sh.crop((0, 0, w - 4, h)), sh.crop((4, 0, w, h)))).mean[0]
print("seam v", round(vert, 1), "h", round(horz, 1), "baseline", round(base, 1), "cost", out["usage"]["cost"])
print("PASS" if max(vert, horz) < base * 1.5 + 2 else "SEAM VISIBLE")
sh.save("offset_view.png")

Reading the numbers

Compare the seam score with the baseline, which is how much neighboring columns differ elsewhere in the image. Busy patterns have a high baseline, so a fixed threshold would fail them wrongly.

If the tile fails

• Ask for a simpler, evenly spaced motif. Large single objects and gradients cross edges badly.
• Open offset_view.png and look. The score finds the seam, but your eyes decide if it matters.
• Fix small seams by cloning: paste a blurred strip over the cross. For big seams, regenerate.
• Try a different model before a different prompt, since the same text gives a different edge behavior on each.

Next steps

Add the check to your prompt regression test so a model swap flags new seams. Once a tile passes, export the widths you need with the srcset recipe, and save only lossless files during repair, per PNG between passes.