Pick the cleanest last frame: sample 24 stills before chaining

When you chain AI clips, the very last frame is often a bad starting still: it can be blurred, mid-gesture or closing on a transition. Sume's video frames API lets you sample up to 24 stills from a clip, so you can pick the cleanest one and send it as the next clip's first_frame.

This replaces the 16 keyframes Luma lists for Ray 3.2, which Sume does not run, with a manual keyframe choice between clips.

Sample, do not just grab the end

Video frames takes one workspace clip and either at[] (1 to 24 explicit seconds) or fps (above 0 and at most 2). It returns durable media.sume.com image artifacts at the source size, as jpeg (default) or png, with an optional max_edge of 16 to 2160. The source must be 300 seconds or shorter.

With fps, Sume uses mid-bin samples: 0.5/fps, 1.5/fps and so on, with a cap of 24 frames. On a 10-second clip, fps: 2 gives 20 stills, and fps: 0.8 over a 30-second clip gives 24 stills with the last at 29.375 seconds.

The request

A submit always returns 202, because the route pins async mode, so do not send mode: sync. Poll GET /v1/video-frames/:id until resource_status is ready. A frame that fails to extract has a url of null and does not fail the job.

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-clip-07" \
  -d '{"video_url":"https://media.sume.com/artifacts/artf_demo/clip07.mp4",
       "fps":0.8,"format":"png"}'

What to look for in the stills

Choose a frame with the subject in a neutral pose, open eyes, no motion blur and the framing you want the next shot to start with. Prefer png for the pick, since it is lossless, and the frame keeps the source size unless you set max_edge.

If you ask for an explicit at of 30 on a 30-second clip, the worker fails with frame_time_out_of_range, because t must be at least 0 and below the duration. Ask for 29.9 instead.

Feed it forward

Send the chosen still to the next POST /v1/videos as a frame_images entry with frame_type: first_frame, on a model whose supported_frame_images lists it. For an extra anchor, also send a last_frame image if the model supports that type. Each frame extraction is billed by its Modal compute (container seconds times Modal list times 1.25, plus the platform fee), and the bill is never more than the hold.

A last-frame chain accumulates drift: every hop is a fresh generation from a still. Keep chains short, and rejoin at hard cuts in Timeline.

Using video inspect instead

If you only need a quick look, video inspect returns a probe plus 8 mid-bin stills by default, at a default max_edge of 768. That is enough to judge a clip but too small for a first frame. Use video frames, which keeps the source size, when you pick a still for the next job.

Video frames allows 1 to 24 values in at[] or an fps up to 2, and the source must be 300 seconds or shorter. A 30-second AI clip is well within that.

A selection routine

• Sample the last 3 seconds with an explicit at list, for example 27, 27.5, 28, 28.5, 29, 29.5 on a 30-second clip.
• Reject frames with motion blur, closed eyes or a half-turned head.
• Prefer a frame the model could plausibly continue from, with space around the subject.
• Send the winner as first_frame and keep the same prompt style for the next clip.

Cost and failure notes

The extract reserves its Modal compute ceiling at submit and bills the actual container time times the Modal list times 1.25, plus the platform fee, never more than the hold. Because a failed instant returns a null url without failing the job, check every frame URL before you pick. For a clip over 90 seconds, warnings[] can include low_confidence_long_video; AI clips under 30 seconds do not hit it.