Test the hardest 5 seconds first: trim, then Recast before the full 30

To test H3 Max Recast on a hard clip before paying for the whole thing, trim the hardest 5 seconds with video trim ($0.02) and recast only that piece. At 768p the test costs $1.875 plus $0.02, or $1.895, against $11.25 for the full 30 seconds. Recast's minimum is 5 seconds, so 5 is the shortest valid test.

What counts as hardest

Pick the stretch most likely to break: two people crossing, a hand over a face, fast motion, a person turning away from the camera, or a head partly out of frame. A test on the easy part tells you nothing about the part that fails.

• People overlap or cross paths.
• A face is in profile or partly hidden.
• The camera moves quickly.
• The person is small in the frame.

The two calls

Trim first. POST /v1/video-trim takes start plus duration and returns a new media.sume.com clip. Use precision: "exact" for a frame-accurate cut. Then send the new URL to Recast as video_url, with one photo per person in reference_image_urls.

curl -X POST https://api.sume.com/v1/video-trim \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: hard-five-trim-001" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/source.mp4",
    "start": 14,
    "duration": 5
  }'

Cost of the test

Recast's Sume rate is the fal list price times 1.25: $0.375 per second at 768p and $0.5625 at 1080p.

Pick the photos before the test

One photo per person, in the order of the people in the clip. A clear, well-lit, front-facing photo works better than a cropped group shot. Use the same photos in the test and the full run, or the test says nothing about the run. For a clip with two people in the hard stretch, the order of reference_image_urls matters; check the result for a swap before you judge anything else, and confirm the clip is at least 5 seconds after the trim.

Recast also takes an optional prompt; leave it empty for the first test and add one only to fix what you see. A test clip is a real job with a real charge, so keep the Idempotency-Key from the request, and reuse it only for a true retry.

What the test proves

It shows whether the photo-to-person match holds in your worst moment, and whether the background and the motion survive. Pull stills from the result with video frames and compare them with the source. It does not prove that the rest of the clip will pass, and it does not prove that the identity will match between chunks if you later join several. If the hard stretch passes at 768p, you have far better odds on the rest; then decide whether to spend the $11.25. If it fails, change the photo, not the settings; a clear, front-facing photo per person is the input that matters.

The full-clip case is in recasting video longer than 30 seconds, and the 768p draft then 1080p final note covers the resolution step.