AI video shows more units than you ship? TikTok Shop's quantity rule

TikTok Shop's AI listing-image policy prohibits showing multiple products when the buyer receives only one or two pieces, and misrepresenting included SKUs, bundles or quantities. Video models like to fill a scene, so check every AI clip: pull stills with Sume's unbilled POST /v1/video-frames and count the units by eye before you post.

The rules come from TikTok Shop's page on AI-generated content for listing images, read 2026-10-02. That page is about listing images; the same logic applies to video, and the platform's wider video rules require the product shown to match what is sold.

Which quantity rules does TikTok Shop spell out?

The AI page groups several rules that a generated clip can break without anyone meaning to.

• Do not misrepresent included SKUs, bundles or quantities.
• Do not show multiple products when the buyer receives only one or two pieces.
• Do not replace the actual product with animals, people or unrelated objects as the main subject.
• Do not show a small item as much larger than its stated size, or use people or furniture to create false scale.
• Product claims must match the listed specifications, such as material, compatibility and volume.

Why do generated clips drift on count?

A video model fills the frame. Asked for a product on a shelf, it can add a second and third bottle behind the first, or duplicate the item mid-motion. Image-to-video from a single photo gives the model the product but not the number, so the count is the model's guess.

Sume cannot tell you the count. It has no object-counting step, so the review is visual: you or a reviewer look at stills and tally.

How do I pull stills to count the units?

Video frames takes one clip on media.sume.com plus either at[] (1 to 24 times in seconds) or fps (above 0 and up to 2, capped at 24 frames), and returns durable image artifacts. It is unbilled. Submit is always 202, and you poll GET /v1/video-frames/{id} until resource_status is ready. If the clip is not on the Sume media host yet, import it first; the media inputs guide explains how.

Use format: "png" for a lossless look, and pick times that cover the start, the middle and just before the end. Every at value must be below the clip duration or the worker fails with frame_time_out_of_range.

import os, time, requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
r = requests.post("https://api.sume.com/v1/video-frames",
    headers={**H, "Idempotency-Key": "count-check-sku-1042"},
    json={"video_url": "https://media.sume.com/artifacts/artf_demo/clip.mp4",
          "at": [0, 2, 4, 5.5], "format": "png"}, timeout=30)
r.raise_for_status()
rid = r.json()["request_id"]
while True:
    g = requests.get(f"https://api.sume.com/v1/video-frames/{rid}", headers=H, timeout=30).json()
    vf = g.get("video_frames", g)
    if vf.get("resource_status") == "ready":
        for f in vf["frames"]:
            print(f["t"], f["url"] or "extract failed")
        break
    time.sleep(3)

How do I write prompts that do not invite extra units?

Say the number. "One bottle, centred, nothing else on the table" leaves less room than "bottles on a shelf". Use singular nouns, and say what the surroundings contain. If you ship a two-piece set, describe exactly two.

Where the count matters most, such as a bundle listing, start from a real photo of the bundle as a reference, then verify the result. Keep the post-check even then: a reference constrains a model, it does not guarantee the number. Because video frames is unbilled, running it on every candidate costs only a minute of looking, so make it a standing step rather than a spot check.

What should I do when a clip fails the count?

Do not post it, and do not fix it with a caption that says what you ship. A truthful caption on a misleading picture is still a misleading picture. Keep the failed stills with your review notes so the next prompt can name the exact problem, such as an extra bottle at the left edge in the second half of the clip. Re-run with a tighter prompt or a different reference, or trim to a segment where the count is right if the whole clip is not needed.

For what happens to videos that break the AI rules, read the post on enforcement and the Creator Health Rating.