Preview a vertical clip fast: video inspect seek fast, then cut exact
Sume video inspect can return stills from the keyframe at or before each time, up to about one GOP early, to skim a long clip before an exact trim.

To skim a long vertical clip quickly, ask Sume video inspect for stills with frames.seek: "fast": each still comes from the keyframe at or before its time, so it can be earlier by up to one GOP (roughly 0 to 5 seconds on typical sources) but never later. Use the default precise seek only when a timestamp must be exact, such as before a trim.
The behavior is in the video inspect docs (read 2026-10-06). Sume says quality, resolution, and transcript do not change with fast; only the decode step is skipped.
What does the frames program allow?
The figures are laid out below.
| Option | Effect | Limit |
|---|---|---|
frames not sent | 8 mid-bin stills (1 fps if the clip is under 8 seconds) | 8 stills |
frames.at[] | Explicit timestamps | 1 to 24 values |
frames.fps | Sample rate, mid-bin | 0 to 2, 24 stills at most |
frames.max_edge | Longest edge of each still | 64 to 2160, default 768 |
How do I skim a 3-minute Short?
A 180-second clip with the default 8 stills gives one still every 22.5 seconds, mid-bin. For a closer look, send frames: {"fps": 0.1333, "seek": "fast"}, which yields about 24 stills across 180 seconds, or one every 7.5 seconds. Each still carries t, url, width, and height, and the grid reports sample_times (what the tiles show) next to requested_times.
When must I switch back to precise?
Before you cut. If you pick a start time from a fast still, the real frame may be up to a keyframe interval earlier, so use precise seek, or trim with precision: "exact" (the default in video trim) and read actual_start_seconds back from the result. A keyframe trim can start a GOP early too, so re-base your times against that field.
Sources
Related posts
More in Media tools
- Reddit's 15-second view rule: a lip-sync clip that stays under it
Reddit's Engaged Video Views beta bills videos over 15 s at 15 s. Sume H3 Max lip sync tops out at 14.8 s of audio, so one clip fits. Specs, math and a caveat.
- Reel captions in two languages: language hint vs auto-detect
How Sume video captions pick a language and a style: the optional language hint, auto-detect, the Hangul rule that returns 400, and a Korean Reel example.
- Reels profile grid crops to 3:4 (1080x1440): check a frame
A Reel cover is center-cropped to 3:4 on the profile grid, per a third-party guide. Pull the frame with Sume video frames and keep faces and titles inside.
- Reference video analysis on Sume: inspect, ingest or video-analyses?
For a reference video, use Sume video inspect for stills and probe, reference ingest for shots and OCR, and note that video-analyses returns 410 on dest.
Written by Sume