Video inspect seek fast vs precise: stills up to one GOP early

Set frames.seek to fast on Sume video inspect when you only want a quick look: each still moves to the keyframe at or before its requested time, so it can show a moment up to one GOP earlier (about 0 to 5 seconds on typical sources) but never a later one. Leave it at precise, the default, when the timestamp has to be accurate.

Both modes return the same quality, resolution and transcript; only the position of the still differs. The behavior is described on the video inspect page.

How the two modes differ

seek lives on the frames object, next to at[] or fps. An object must carry exactly one of at[] (1 to 24 values) or fps (above 0, at most 2, mid-bin, limited to 24 stills).

Reading the response

With fast, each grid reports seek: "fast". sample_times are the instants that the tiles actually show, and requested_times are the instants in your program. Compare the two before you quote a timestamp to anyone.

Example: you ask for at: [60] on a clip with a keyframe every 4 seconds starting at 0. The nearest keyframe at or before 60 is 60 itself, so nothing moves; ask for 62 and the still comes from 60. With keyframes every 5 seconds, a request at 64.9 shows 60.

{
  "video_url": "https://media.sume.com/artifacts/artf_demo/talk.mp4",
  "frames": { "fps": 0.2, "seek": "fast", "max_edge": 512 }
}

Limits that still apply

The source must be 1800 seconds or shorter, and a call returns at most 24 stills. The image edge defaults to 768 and accepts 64 to 2160. The call must use a media.sume.com clip your workspace already owns; import other files first. Sume bills probe and stills by their Modal compute, so there is no fixed per-call price to quote here.

If you need the exact frame at the source size, video frames is the better route, with its own 16 to 2160 edge clamp.

When to use precise anyway

Pick precise for anything a person will quote: a transcript check, a caption timing review or a frame you will paste into a ticket. The cost of precise is decoding time, which is why the docs offer fast only for rough looks.

A useful split is a two-pass review. First pass: fps: 0.2 with seek: fast across a 10-minute file gives 0.2 x 600 = 120 candidates, which the call caps at 24, so lower the rate to 0.04 to get 24 stills spread across the file. Second pass: use precise with at[] set to the few instants you found interesting.

• 24 stills is the ceiling per call, for both at[] and fps.
• fps above 2 is rejected; fps below the ceiling spreads the stills evenly.
• Transcription is a separate flag billed at the STT rate of $0.01 per audio minute.