Review a batch of Omni drafts fast: video-inspect stills, then re-run

To review many Gemini Omni drafts cheaply, generate them at 360p on Sume, run a video_inspect call on each with seek: fast and a low frame count, look at the stills, and re-run only the winners at 1080p. Fast seeking snaps each still to the nearest earlier keyframe, so it is for skimming; use precise seeking only when the exact instant matters.

Why draft at 360p?

The Video Router docs list gemini-omni-flash-1.1 at 360p, 720p, 1080p and 4K, 3 to 10 seconds, and bill provider list times 1.25 per output second by resolution. Google's pricing page says Omni video is billed at 5,792 tokens per second of 720p video, which is about $0.10 a second; Google's Omni page says 1080p and 4K are upscaled.

I could not find a published per-token figure for 360p on those pages, so I do not quote a 360p price. Check the live rate with GET /v1/video-router/models before sizing a batch.

How does fast seek help?

Video inspect samples stills from a clip you already host on media.sume.com. By default it returns 8 mid-bin stills. With frames: { fps: n } or at[] you choose the sampling, 1 to 24 stills per call, and max_edge defaults to 768.

seek: "fast" skips the exact decode and snaps each still to the keyframe at or before its instant, by up to one GOP, roughly 0 to 5 seconds on typical sources, never later. Quality and resolution are unchanged. The sample_times field reports the instants the tiles really show, so the mismatch is visible, not hidden. A 3 to 10 second Omni clip has few keyframes, so use precise when your stills must be evenly spaced.

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-draft-007" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/draft-007.mp4",
    "frames": { "at": [0.5, 3, 6, 9], "seek": "fast", "max_edge": 512 }
  }'

What is the full loop?

Submit every draft with its own Idempotency-Key and a prompt id you keep in a table. When each job finishes, inspect its result. Review the stills, mark the drafts you like, and re-submit those prompts at 1080p.

Sume does not rank drafts for you. A re-run is a new generation, not an upscale of the draft, and Google's Omni page lists no seed field among its settings, so the 1080p run can differ from the 360p draft. Treat the draft as a screening step, not a preview of the final frame.

What can go wrong?

A clip whose motion happens between keyframes can look static in fast stills. If two drafts look identical in the tiles, re-inspect one with precise. Also check probe.has_audio on the first call, since Omni's audio is always on and a silent result would be unexpected. For waiting on many jobs see Jobs and results.