Did my Wan 3.0 clip actually cut between shots? Check with stills

To check whether a Wan 3.0 clip really cuts between shots, pull stills at the seconds where your prompt promised a cut and compare them. On Sume that is POST /v1/video-inspect with frames.at: up to 24 timestamps, one image each. Alibaba's Wan 3.0 README (read 2026-10-05) lists auto scene splitting among the model's features, which means the model decides where to cut when a prompt describes several scenes. It is a model behavior, not a request flag: Sume's catalog has no parameter that turns it on, so verification is the only control you have.

Why check at all

A multi-scene prompt can come back as one long take that morphs between scenes, or as a clean cut with the scene you asked for, or as a cut a second early. If you plan captions, music beats or a Timeline join around the cut times, a one-second drift matters. Two stills straddling each planned cut tell you which of the three outcomes you got.

Sample the clip

Import the finished clip so it is a media.sume.com artifact, since video inspect does not fetch from the open internet, then request stills just before and after each planned cut. The default is eight mid-bin stills; explicit at[] gives you control, and max_edge defaults to 768. For a 27 second clip with cuts planned at 9 and 18 seconds:

curl -X POST https://api.sume.com/v1/video-inspect \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: inspect-cuts-001" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/wan-clip.mp4",
    "frames": {"at": [8.0, 10.0, 17.0, 19.0], "seek": "precise"}
  }'

Which tool answers which question

Video inspect returns probe facts and stills, not a list of cuts. The docs say plainly that it returns no typed scenes. If you need cut times rather than pictures, reference ingest answers with shots[], frame-exact cuts voted by ffmpeg scdet and PySceneDetect, plus keyframes. It is not available everywhere: the page says it is listed only where the feature flag allows it (development on, production opt-in), it reads one clip of at most 300 seconds, and it needs an Idempotency-Key.

What to do with the answer

If the cut is where you wanted it, move on to captions or a Timeline join. If the model blended two scenes, split the job: one wan-3.0 request per scene, reusing the same references, then join them with Timeline 1.0. That costs more requests but fixes the cut times, because each scene is its own file.

If the cut is a second off, you can keep the clip and trim it with video trim, which cuts a [start, end) range into a new MP4 without changing the source. Record the planned and actual cut times in your notes for the next prompt: if the model consistently cuts early, say so in the prompt by padding the first scene.

One more check is worth the extra still: look at the last frame. A clip that ends mid-motion will not join cleanly to the next slot in a timeline, and a clip whose final scene has changed look compared with the start is a sign the model drifted. Ask for stills at the very start and the last second, and compare them against the references you sent.