January stocktake sale video: one caption cue per price

For a January stocktake sale video, put each price in its own caption cue instead of painting it into the product photos. Render the product stills as a 15-second silent clip in Timeline 1.0 ($0.10 per output minute), then call POST /v1/video-captions with five cues, each 3 seconds long, for $0.20. The finishing cost is $0.30, and when a price changes mid-sale you re-run only the $0.20 caption call, not the render.

Clearance prices are the text most likely to change: stock runs out, a discount deepens, the sale gets extended. Keeping them out of the images is the whole idea.

Why cues instead of text inside the image?

Image models can garble digits and percent signs, and an edited price means a new image and a new render. The captions page says authored overlay text goes in cues (or segments) with text, start and end, burned without speech-to-text. A silent product clip has no speech anyway, and a caption job on it fails with caption_no_speech unless you send cues.

How long should each price stay on screen?

Three seconds is a practical window for a short line such as a product group and a discount. Five prices fit in 15 seconds, which matches five 3-second still slots in the render. Keep the cue windows aligned with the slots, so the price you read matches the product on screen.

{
  "cues": [
    { "text": "Winter coats 30% off", "start": 0, "end": 3 },
    { "text": "Boots 25% off", "start": 3, "end": 6 },
    { "text": "Scarves 2 for 1", "start": 6, "end": 9 },
    { "text": "Gloves 20% off", "start": 9, "end": 12 },
    { "text": "Stocktake sale ends Sunday", "start": 12, "end": 15 }
  ]
}

How do I build the base clip?

Import five product photos so they are Sume-hosted, then create a Timeline 1.0 document with audio.mode set to silence, audio.duration_seconds of 15, and one video[] slot per photo (start 0, 3, 6, 9, 12 and duration 3). Run POST /v1/timeline-1.0/plan first; it is unbilled and shows the estimate. Use fit: cover for photos that are already vertical and contain for ones that are not.

Then send the render with an Idempotency-Key, poll the job, and take the video_url from the result.

What does the sale cost to make?

Rates are from the Sume docs, which advise confirming live rates in GET /v1/catalog.

What can go wrong?

Cue windows that overlap are hard to read, so keep each end at or before the next start. Keep the text short enough to fit the frame, and check the clip on a phone before you post.