Multi-shot AI video prompts: MiniMax H3 shot labels and timecodes

To get several shots in one MiniMax H3 clip, write them into the prompt: MiniMax's prompt guide labels shots [Shot 1], [Shot 2] and gives later shots a timestamp such as "At 00:03.500, the camera cuts to...", while fal's guide uses timecoded blocks like [0–2 seconds]. The model does multi-shot natively, and Sume has no separate shots field: the storyboard is prompt text.

Syntax is from MiniMax's prompt guide and fal's guide, and MiniMax's announcement for the native multi-shot claim; request fields from the Sume Video generation docs, read 2026-09-29.

What does a multi-shot prompt look like?

Keep to the length you buy: duration is 5–15 seconds on Sume, so timecodes past your duration have nothing to land on. The shot order below adds up to 10 seconds.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: h3-shots-001" \
  -d '{
    "model": "minimax-h3",
    "prompt": "[0-3 seconds] Wide shot of a lighthouse at dusk, waves breaking. [3-6 seconds] The camera cuts to a close-up of the keeper lighting the lamp. [6-10 seconds] Pull back outside as the beam sweeps across the water. Sound: wind, gulls, a low cello.",
    "resolution": "768p",
    "aspect_ratio": "16:9",
    "duration": 10
  }'

Which syntax should I use?

What keeps a multi-shot clip coherent?

MiniMax's guide says to establish the style at the start of Shot 1, keep the character consistent and describe observable changes. fal's advice is to describe transitions as physical events rather than named effects. If you have character images, send them as references and name them in the prompt so each shot reuses the same look.

Does Sume validate the shots?

No. Sume checks the request envelope, such as duration, resolution and aspect ratio, and sends the prompt to the provider. Whether shot 2 lands at 3 seconds is the model's behavior, so check the output.