Instagram Align for Reels: match AI shots with the last frame

Instagram's Align tool helps you line up objects between two clips you record in the app, and Sume has no equivalent button. The AI-video version of the same idea is mechanical: pull a still from near the end of shot A, then send it as the first frame of shot B so the next clip starts where the last one stopped.

The Align line comes from Instagram's Reels feature page (read 2026-10-10): "Create seamless transition moments with Align to line up objects between clips. Set a timer to record hands-free, and adjust speed for video and audio." The page says nothing about AI clips, so everything below about Sume is Sume's own documented behavior, used to get a similar result.

The workflow in four steps

Each step uses a surface the Sume docs describe. Nothing here needs a model that Sume does not list.

• Generate shot A with an id that accepts frame_images or an image_url, then import it into your workspace, because the Video frames API only reads a media.sume.com clip your workspace already owns.
• Extract a still with POST /v1/video-frames and at: [t], where t is a time just inside the end of the clip. The docs say frames come back as durable media.sume.com images at source size, and that a frame that fails for one instant returns url: null without failing the job.
• Send that still as the first frame of shot B. On /v1/videos, the frame_images array takes frame_type: first_frame. On the Video Router, Gemini Omni Flash 1.1 takes image_url for image-to-video.
• Join A and B in Timeline 1.0, which assembles ordered video slots plus one audio spine into a single MP4.

Which models take a first frame

Not every id accepts a start image, so read the models endpoint first. The Video Generation page says each model reports supported_frame_images, and gives the Seedance 2.0 entry as first_frame and last_frame. The Video Router page lists image-to-video for Gemini Omni Flash 1.1, with an optional end_image_url.

Using last_frame for a loop or a match cut

Where a model lists last_frame, you can also pin where shot B ends. If shot B ends on the same still as shot A began, the pair loops. If it ends on a new still you chose, you control the destination of the transition rather than leaving it to the model. Pin the first and last frame, and describe only the motion in between in the prompt.

Keep shots short. Instagram's page says you can create multi-clip videos up to 3 minutes in the app, and a chain of 3 to 10 second AI shots reaches that length only with many links; each link adds a chance of drift from the original look. Re-anchor every few shots by sending a fresh reference image rather than relying on the last still alone.

Check the seam, not just the clip

Align exists because an unmatched cut reads as a jump. Review the seam at normal speed and at the speed Instagram's own tool offers, since the same page mentions adjusting speed for video and audio. If the objects do not line up, extract another still a little earlier in shot A and regenerate shot B with a new idempotency key; the docs say reusing a key returns the original job.

One caveat on inputs: the Video Generation troubleshooting section says reference images must be available over public HTTPS. Confirm that the still URL you send meets that rule before you submit a batch, since a failed submit wastes a round trip.