Coffee or tea pouch photos with AI: zip top, crimp and print

A coffee or tea pouch has a few shape details that identify it: the flat crimp on top, a tear notch, a resealable zip, maybe a clear window. Generated scenes often smooth them away and produce a generic bag.

Using Sume's Image API, you give the pouch photo as an input_references item and ask for a morning-kitchen scene around it.

List the shape features

Prompts that say "keep the pouch" are vague. List the crimp, the zip line and the window by name, and say the pouch stands upright on its own.

import os, requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
resp = requests.post("https://api.sume.com/v1/images", headers=H, timeout=90, json={
    "model": "openai/gpt-image-2.5",
    "prompt": "Place this exact coffee pouch standing upright on a kitchen counter next to a ceramic mug, steam rising, morning light. Keep the flat crimped top, the zip line, the window and the front print as in the reference.",
    "input_references": [
        {"type": "image_url", "image_url": {"url": "https://example.com/pouch.jpg"}}
    ],
    "aspect_ratio": "auto",
})
resp.raise_for_status()
if resp.status_code == 202:
    raise SystemExit("job envelope: poll data.status_url")
body = resp.json()
print([d["url"] for d in body["data"]], body.get("usage"))

Print and seal

Front print is a smaller version of the label problem: a name and a roast level can drift. If exact print matters, use the approach in the label compositing post and paste the real front panel back. For seal and shape, compare side by side with the reference:

• The crimp is flat and straight, not wavy.
• The zip line sits below the crimp, at the same distance.
• The window, if present, shows coffee beans or leaves, not a blank.
• The pouch stands without a visible prop.

Run one image per scene, review them together and keep the passing ones. The cost of each call is in usage.cost, which is enough to set a per-SKU budget before you generate a season of scenes.

Before you run a catalog

Run one SKU end to end first. POST /v1/images returns the images directly when it finishes within 30 seconds; past that it returns a 202 envelope with status_url and result_url, which the jobs and results guide explains. Handle that branch before you loop over a catalog, and write each result's URL and usage.cost to a file keyed by SKU, so a failed run restarts where it stopped and nothing is paid for twice.