AI virtual try-on looks fake? A QC checklist for print and hands

Compare the generated image with your real garment photo at full size, in the same order every time: colour, print placement, neckline and sleeve length, hands, then the background. For a clip, pull still frames at fixed times with Sume's POST /v1/video-frames and run the same checks on frames 0, middle and end, because a try-on can be right on frame one and wrong by the last second.

ChatGPT Try On, launched October 1 2026, is reported to carry the warning that images may not reproduce the person or product exactly. If the biggest AI shopping assistant says that, a seller publishing its own try-on should check, too.

What does the vendor say can go wrong?

Retail Technology Innovation Hub reports that OpenAI notes try-on images may not reproduce the person or product exactly and do not guarantee size or fit. That is a reported caveat, not OpenAI's own page text, which we could not open. It matches what the output looks like in practice: the failures are small and specific, which is why a checklist works better than a general look.

What should you check, in what order?

Start with the facts the shopper can verify against your listing, then move to the rendering. The first four are cheap to check against your packshot; the last three need a closer look at full resolution.

How do you check a video frame by frame?

Sume's Video frames takes one clip hosted on media.sume.com plus either an at[] list of seconds (1 to 24 values) or an fps, and returns durable images at source size. png is the lossless option for inspection. Every at value must be inside the clip's duration, or the job fails with frame_time_out_of_range. The submit returns 202; poll the resource until resource_status is ready.

Pull the first frame, the middle and the last, then lay them beside your packshot.

import os
import time
import requests

key = os.environ["SUME_API_KEY"]
h = {"Authorization": f"Bearer {key}"}
body = {
    "video_url": "https://media.sume.com/artifacts/artf_demo/tryon.mp4",
    "at": [0, 1.5, 3, 5],
    "format": "png",
}
r = requests.post("https://api.sume.com/v1/video-frames", headers={**h, "Idempotency-Key": "qc-tryon-1042-v1"}, json=body, timeout=60)
r.raise_for_status()
fid = r.json()["data"]["video_frames_id"]
while True:
    time.sleep(5)
    g = requests.get(f"https://api.sume.com/v1/video-frames/{fid}", headers=h, timeout=60).json()["data"]
    if g["resource_status"] == "ready":
        for f in g["frames"]:
            print(f["t"], f["url"])
        break
    if g["resource_status"] in ("failed", "canceled", "archived"):
        raise SystemExit(g.get("error") or g["resource_status"])

Who should do the review?

Someone who has handled the real garment. A designer or a stylist notices a wrong seam or a flattened collar faster than an engineer does, and the check takes under a minute per image once the packshot is open beside it. Make the person who approves the image also the person who owns the size chart, so the two never disagree on a published page.

Automate only the plumbing: the extraction of frames, the storing of the result under the SKU, and the side-by-side sheet. Leave the yes or no to a human, and record it.

What do you do when a check fails?

Change one thing at a time. For a still, tighten the prompt on the detail that failed and rerun; for a mask edit on a colour problem, see the garment colour post. For a Format run, continue it with previous_run_id and name only the scene that failed, as in the retry one scene post, so you do not pay for the whole video again. If the garment itself is the problem, a cleaner packshot is a better fix than a longer prompt.

Keep your review notes next to the SKU: which check failed, and what you changed. They become your prompt library.

• Review at 100 percent, not on a phone-sized preview.
• Always look at the last frame of a clip, not just the first.
• Re-run one change at a time so you know what worked.
• Never publish a size or fit claim taken from a generated image.
• Label AI-generated try-on media where a platform asks.