Chain Wan 3.0 clips: last frame of one is the first frame of the next
Extend a Wan 3.0 scene past one clip by extracting a still near the end with video-frames and passing it as the first frame of the next request.
To continue a scene past one Wan 3.0 clip, extract a still from just before the clip ends with POST /v1/video-frames, then send that image as image_url (the first frame) on the next wan-3.0 request. Wan 3.0 accepts first and last frames and 2-30 s per clip, so one clip can already be long; chaining is for scenes that must continue after 30 s or that you want to steer clip by clip.
Steps follow Sume's Video frames docs and Video Router docs, read 2026-10-02. Test one hop at 480p before you commit to a long chain.
How do I pull the last frame?
Video frames takes one media.sume.com clip and either an at list of seconds or an fps value, and returns image artifacts at source size; extraction is unbilled. Each at value must satisfy 0 <= t < duration, otherwise the worker fails with frame_time_out_of_range, so ask for a time just before the end, such as the clip length minus 0.1.
Read source_duration_seconds from the result to confirm what the worker probed. A single instant that failed comes back with url: null and does not fail the job.
| Field | Rule |
|---|---|
video_url | This workspace's media.sume.com artifact |
at[] | 1-24 seconds, each at least 0 and less than the duration |
fps | Greater than 0 and at most 2; capped at 24 frames |
format | jpeg (default) or png |
| Billing | Unbilled |
How do I feed it to the next clip?
Send the frame URL as image_url and write a prompt that continues the action. The Video Router docs list Wan 3.0 as accepting text, image-plus-end-frame and reference inputs. Input image URLs must be fetchable public HTTPS URLs, so if the extracted URL is refused, re-host the still and send that.
Use a fresh Idempotency-Key for each hop and keep the first clip's job id in your records so a failed hop can be retried without redoing earlier ones.
What goes wrong with chained clips?
Drift. Each hop starts from a compressed still, so faces, logos and colors can shift over several hops. Check a contact sheet from video frames before you stitch. Join the clips afterwards with Timeline 1.0, which sequences up to 200 video slots with transitions.
If one 30-second Wan clip covers the scene, skip the chain: see 30-second single take versus stitched clips on Wan 3.0 and getting the last frame of a video.
Sources
Related posts
More in Use cases
- Wayfair image guide: five shot types, 1000x1000 minimum
Wayfair wants product images of at least 1,000 x 1,000 px in five shot types. Plan a silo, two environmental, functional and dimensional set from one packshot.
- WCAG 1.2.5 audio description: narrate a video with TTS timestamps
WCAG 1.2.5 asks for audio description on prerecorded video. Build the narration track with Sume TTS word timings and timeline audio concat.
- Webinar highlight reel: trim five moments, stitch, add cue captions
Cut a 60-second webinar highlight reel from a recording: five trims at $0.02, one Timeline render at $0.10, and caption cues at $0.20, about $0.40 in total.
- Webinar recording to shorts: an API pipeline with Sume
Turn a webinar recording into short clips with four calls: inspect and transcribe by sentence, pick moments, trim, caption. Limits and costs included.
Written by Sume