Reference ingest for a 30-second ad: shots, text and audio in one read
Reference ingest reads one clip of up to 300 s into a manifest of shots, OCR text and audio facts. Use it before remaking a winning 30-second ad.
How do you break a competitor's winning 30-second ad into something you can remake? Read it once with reference ingest: it returns shot cuts, on-screen text, audio facts and an overview strip as a manifest, so your brief starts from measured facts instead of a viewer's memory. The clip must be one your workspace already owns, up to 300 seconds long.
The timing is right for 30-second work. ByteDance says Seedance 2.5 generates up to 30 seconds per pass with joint audio and video, and accepts up to 30 images, 10 video clips and 10 audio clips as references. Sume's catalog lists seedance-2.5 at 4-30 seconds, so a 30-second reference-driven remake can be one job.
What the manifest holds
POST /v1/reference-ingest takes a video_url (a media.sume.com artifact, asset or chat attachment of your workspace). Idempotency-Key is required. The default mode is sync: the handler waits up to 30 seconds and answers 200, or answers 202 with a queued job.
| Field | What you get |
|---|---|
| shots[] | Frame-exact cuts, one sharp keyframe per shot, palette, luma, motion class |
| text_tracks[] | OCR lines at source resolution with a box, span, confidence and role_hint |
| audio | Loudness gate (silent or not), speech presence, beats for music, optional transcript |
| boundaries[] | Mechanical visual-change candidates beside the shots |
| overview | One labeled strip, up to six tiles |
| uncertain[] | The only reasons to look at a frame again |
Remaking it
Read the shot list and the on-screen text, then decide which parts are structure (hook, proof, offer) and which are content you must replace. Your product images and your voice replace the original's; the pacing stays.
If audio.silent is true, plan new music and discard the source audio. Never request speech-to-text on a silent track: the transcript is billed only when the track is not silent and speech is found, and otherwise the charge settles to zero.
Know the limits
Clips over 300 s fail with source_too_long_for_reference_ingest. The request has no filtergraph or codec fields; those keys return 400 ffmpeg_fields_rejected. Sume bills the manifest by its media-compute usage, with the reservation settling to what ran.
Availability of this route is gated per environment, so confirm it works for your key before building a pipeline on it. Only analyze clips you have the right to use as a reference, and replace the original's people, music and branding in the remake.
Reading the manifest well
The labeled strip is handy for review because it shows the frames the shot list refers to. Use the shot boundaries to count the cuts and the average shot length, then match that rhythm in your remake. Use the OCR text to see which on-screen claims the original made, and rewrite them for your own product and evidence.
Sources
Related posts
More in Use cases
- Fix one bad shot in a 3-minute AI video for $2.18
A 12-shot, 3-minute AI video costs $22.86 to build. Re-rendering one 15-second shot and re-composing costs $2.18 on Sume.
- Remove a date stamp from a scanned photo: crop first, then a mask edit
Remove an orange date stamp from a scanned family photo: crop it for free in Pillow, or mask it and edit with openai/gpt-image-2.5 from $0.0094 an image.
- Remove a person or object from a finished clip with Gemini Omni edit
Delete a passer-by, a sign or a cable with one video_url request to Gemini Omni Flash 1.1. There is no mask field, so the prompt does the work. Cost per clip.
- Remove a price sticker from a product photo and check the label
Remove a sticker from a product photo with Ideogram 4.5 on Sume, then check the label text underneath. A small mask on GPT Image 2.5 limits the edit.
Written by Sume