Put a promo code on screen in an AI product video with caption cues

To put a promo code on screen in an AI-generated product video, do not ask the video model to draw the letters. Generated clips garble small text, and a wrong code costs you a sale. Instead, render the clip silent, then burn the code with the Sume video captions API using authored cues, which place text you wrote at times you choose, with no speech recognition involved. The job costs $0.20 for a video of 60 seconds or less (read 2026-10-05).

Why cues, not a transcript

A silent image-to-video clip has no speech, so a plain caption job fails with caption_no_speech. The docs say to send cues (or segments) with text, start and end to burn authored overlay text without ASR. Because the text is yours, the code string is exact every time.

Meta says Reels default to sound on, so a spoken code is also reasonable, but Meta's Reels ads page pairs 9:16 video with key messages inside the safe zone. On-screen text covers viewers who scroll with sound off. Set design.placement.anchor_ratio to move the line away from platform buttons.

A runnable helper

The helper below posts one caption job per SKU. The Idempotency-Key combines SKU and code, so retrying the same pair replays the first job instead of billing again, while a new code on the same SKU is a new job.

import json, os, urllib.request

KEY = os.environ["SUME_API_KEY"]


def burn_code(video_url, sku, code):
    body = {"video_url": video_url, "style": "punch", "cues": [
        {"text": "BLACK FRIDAY", "start": 0.0, "end": 2.0},
        {"text": "Code " + code + " at checkout", "start": 2.0, "end": 4.5},
    ]}
    req = urllib.request.Request(
        "https://api.sume.com/v1/video-captions",
        data=json.dumps(body).encode(),
        headers={"Authorization": "Bearer " + KEY,
                 "Content-Type": "application/json",
                 "Idempotency-Key": "promo-" + sku + "-" + code})
    with urllib.request.urlopen(req) as resp:
        return json.load(resp)


if __name__ == "__main__":
    url = "https://media.sume.com/artifacts/demo/clip.mp4"
    print(burn_code(url, "sku1", "BF26"))

Check the pixels, not the request

After the job finishes, extract two stills with video frames: POST /v1/video-frames with at: [3.0], one frame per cue window. Look at the stills, or run your own text check against the known code. Video frames is billed by Modal compute, not a fixed price, so treat it as a small line and read the actual usage afterwards.

Keep the ad text separate from the burned-in cues: the TikTok ad specifications say the ad captions of non-Spark ads are white in a fixed font and carry no clickable links, @ or hashtags, so put the code in the video and not only in the ad text. At $0.20 a job, 100 SKUs cost $20.00 for one code; changing the code means paying the caption line again, so decide the code before the batch, as the 100-ad finishing budget shows.