Serum label close-up: three crop punch-ins from one clip for $0.06

To pull three close-ups out of one serum clip, send the clip to POST /v1/video-filter three times with a single crop op each: fractions of the frame for x, y, width and height. Each job returns a new MP4 for $0.02, so three punch-ins cost $0.06, and the check endpoint that validates a crop first is free.

A serum is a small, label-heavy product. The wide shot sells the mood, and the close-ups on the label, the dropper and the glass do the information work. You can get them from one good take instead of three, which matters when the shoot is a single morning.

Fractions, not pixels

A crop op is {op: crop, x, y, width, height}, all fractions of the source frame. x and y are in 0 to 1, width and height are in 0.05 to 1, and x + width and y + height must each be at most 1; otherwise you get video_filter_crop_out_of_bounds. The compiler rounds the result to even dimensions for yuv420p. The output keeps the source's frame rate and audio unless the program changes them.

The crop is a plain rectangle, not a camera move, so it returns a smaller picture of that region. When you place it in a Timeline slot, the default fit: cover fills the 1080 by 1920 frame, which is what turns the rectangle into a punch-in.

Check before you pay

These rectangles are examples, not measurements of any footage. Pull one still frame of your own clip, mark the label and dropper, and divide by the frame size to get your fractions. A crop that is narrow in a 4K source still has plenty of pixels; a crop of 0.4 of a 1080p frame is 432 by about 486 pixels, and that will look soft when cover-fit onto a 1080 by 1920 slot, so prefer it for a cutaway of a second or two.

The free check validates the program first. It runs the schema, the op whitelist and the host preflight, and returns valid, diagnostics and the compiled filter names with an estimate. A program that passes can still fail on the worker for memory or time, and then you get a structured job error.

Source clips are limited to 300 seconds, so a long product film should be trimmed first with video trim at $0.02.

Pick fractions that keep the aspect of your delivery frame in mind. A vertical slot of 9:16 is tall, so if you crop a wide rectangle, cover-fit will trim its sides again, and a label at the edge can be cut off. A tall rectangle survives the fit better. Leave a margin of 0.03 to 0.05 around the thing you are showing.

Because each job reads the same original, you can run the three crops at once; none depends on another.

Each result is a new artifact with its own video_url; the source never changes, so you can come back later and cut a fourth close-up from the same take without redoing the first three. Budget by crop count: a 20-SKU serum range with three close-ups each is 60 jobs, or $1.20 on the same flat rate.

The check, then three jobs

import os
import uuid
import requests

API = "https://api.sume.com/v1/video-filter"
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
src = os.environ["SERUM_URL"]  # media.sume.com clip
crops = [(0.30, 0.25, 0.40, 0.45), (0.55, 0.05, 0.35, 0.40),
         (0.10, 0.55, 0.40, 0.40)]
for n, (x, y, w, h) in enumerate(crops):
    body = {"video_url": src,
            "ops": [{"op": "crop", "x": x, "y": y, "width": w, "height": h}]}
    chk = requests.post(f"{API}/check", headers=H, json=body, timeout=60).json()
    print(n, chk.get("valid"), chk.get("next_action"))
    if chk.get("valid"):
        r = requests.post(API, json=body, timeout=60,
                          headers={**H, "Idempotency-Key": f"serum-{n}-{uuid.uuid4()}"})
        print(r.status_code, r.json().get("request_id"))

Assemble

Put the three results into a Timeline render with the wide shot, in the order wide, label, dropper, base, wide. Five slots of 2 to 3 seconds is a 12-second spot, one output minute at $0.10, so the whole edit costs $0.16.

Keep what the frame says conservative. A close-up of a label is accurate product description; a caption that says what the serum will do to skin is a claim, and nothing in the API checks it for you. For the same idea on an AI-generated clip see the second-camera punch-in post.