Text edited in an ad image: a dimension check before you ship

After an AI edit that replaces text in an ad image, check three things before it ships: the output has the same pixel dimensions as the source, the aspect ratio matches, and the new text reads correctly. The first two take a few lines of Pillow; the third needs eyes.

Higgsfield lists Ideogram 4.5 reference edits and text replacement that preserve layout (Higgsfield changelog, read 2026-10-02). Whatever model runs the edit, verify the claim on your own file.

Why can an edit change the shape?

Sume's Image API docs say that on edit and image-to-image calls you should prefer aspect_ratio: "auto" to match the reference, and that omitting the field is not the same as auto. An edit sent without it can come back in a different shape from the source.

How do I check dimensions in Python?

Download the source and the result, open both with Pillow, and compare. Allow a tiny tolerance only if your pipeline resizes on purpose; otherwise assert equality.

from PIL import Image


def check_edit(source_path: str, result_path: str) -> None:
    src = Image.open(source_path)
    out = Image.open(result_path)
    if src.size != out.size:
        raise SystemExit(f"size changed: {src.size} -> {out.size}")
    src_ratio = src.width / src.height
    out_ratio = out.width / out.height
    if abs(src_ratio - out_ratio) > 0.001:
        raise SystemExit("aspect ratio changed")
    print("dimensions ok", out.size)


check_edit("banner-us.png", "banner-de.png")

What else belongs on the checklist?

Dimensions are the machine check. The rest is a short human pass per market.

What if a check fails?

Resend the edit with aspect_ratio: "auto", a prompt that names the exact old and new text, and the words "change nothing else". If a small region keeps drifting, a masked edit with mask_url limits the change to that region on ChatGPT Image 2.5. The request fields are in the Image API docs.