Check a restyled video kept the camera move: pull matching frames

To check that a restyled video kept the source's camera move, extract stills at the same timestamps from both clips and compare them side by side: the framing, the subject's position, and the horizon should line up even though the look changed. Sume's POST /v1/video-frames does the extraction, takes at[] times, returns durable image artifacts, and is unbilled. Higgsfield's changelog says Genjutsu Restyle keeps motion and camera; this is a way to verify that kind of claim yourself on any restyle tool.

The endpoint facts are from Sume's Video frames page, read 2026-10-02.

How do I extract the frames?

Send one video_url and exactly one of at[] or fps. The clip must already be a media.sume.com artifact or asset in your workspace; there is no open-internet fetch, so import it first with POST /v1/media-imports. Submit returns 202; poll the GET or GET /v1/jobs/:id/status. format is jpeg by default or png, and max_edge clamps the long side from 16 to 2160.

curl -X POST https://api.sume.com/v1/video-frames \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: frames-source-001" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/source.mp4",
    "at": [0, 1, 2, 3]
  }'

What should I compare?

Use the same times on both clips: first frame, a frame mid-move, and the last frame. Check subject position, camera angle, and where cuts fall. A restyle that keeps a look but shifts a cut by a second is a motion change, even if every single frame looks good.

Is it billed?

The docs say video frames is unbilled, the same class as a screenshot hop: no seat and no reservation. That makes a four-frame check on every restyle cheap enough to be routine.

Where does this fit?

Run it after the restyle and before you publish. Three ways to restyle a clip on Sume lists the routes it applies to.

For a longer clip, use fps instead of a list of times: a low rate such as one frame per second gives a contact sheet of the whole move. Put the two sheets side by side and look for the frame where the camera starts and stops moving. If the output reaches that point earlier or later than the source, the timing changed even though each look is clean.