Swap the person in a clip: reference video or H3 Max Recast on Sume
Seedance 2.5 offers reference-based editing. For a straight person swap, Sume lists H3 Max Recast. When to use which, and the inputs each needs.

If you want the same shot with a different person in it, the Sume route built for that is h3-max-recast: it replaces the people in a source video with people from reference photos. Seedance 2.5 lists reference-based editing (ByteDance Seed, read 2026-10-05), and Sume's video docs say Seedance 2.x accepts video references, which gives you a looser, regenerate-from-reference route. Recast keeps the source performance; the Seedance route re-imagines the shot.
Inputs and limits
The Recast docs state that duration is the source length, so you do not choose it. The Seedance route lets you choose duration but gives up exact fidelity to the source performance.
| Item | `h3-max-recast` | `seedance-2.5` with a video reference |
|---|---|---|
| What it does | Replaces each person in video_url with a person from a photo | Generates a new clip guided by your reference video and images |
| Source video | video_url, 5 to 30 s | Reference video(s) in input_references |
| Photos | 1 to 4 reference_image_urls, one photo per person | Image references; read the catalog for the count |
| Resolution | 768p or 1080p | 480p, 720p or 1080p |
| Prompt | Optional | Describes the whole clip |
| Duration | The source length | 4 to 30 s, your choice |
| Source performance kept | Yes, by design | No; a new take guided by the reference |
Choose by what must stay fixed
- The action, timing and camera of the original must not change:
h3-max-recast. - Only the idea of the shot matters and you want a fresh take: Seedance 2.5 with a reference video.
- The source has more than four people: Recast takes 1 to 4 photos; trim or crop the source to the people who change.
- The shot is over 30 seconds: split it with video trim first.
A Recast request
Photos must be one per person. The prompt is optional; leave it out unless you need to describe the swap.
curl -X POST https://api.sume.com/v1/video-router/generate \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: recast-001" \
-d '{
"model": "h3-max-recast",
"video_url": "https://example.com/source.mp4",
"reference_image_urls": ["https://example.com/new-person.png"],
"resolution": "768p",
"duration": 12,
"mode": "async"
}'Consent and checks
A person swap changes how someone appears on screen. Use only people who have agreed, keep their consent with your project files, and label the output where a platform or law requires it. Do not use photos of people from unknown sources.
After the job, check three moments with video frames: the first second, the middle and the last. Look for identity drift, hands, and any frame where the original person is visible. If a frame shows the original, fix the source crop or re-run.
- Keep consent records.
- Check first, middle and last seconds.
- Confirm no frames show the original person.
Photo preparation
Use one clear, front-facing photo per person, well lit, without sunglasses or heavy filters. The Recast route takes one photo for each person it replaces, so order the photos to match the people in the source, left to right, and describe the mapping in your notes. If the source has four or fewer people but only two should change, crop or trim the source to those two.
Only use photos of people who have agreed to appear. Check the clip afterward with a few stills from video frames, comparing the first, middle and last seconds.
Sources
Related posts
More in Use cases
- Replace audio from 12.4 s to 14 s of a clip with Timeline parts
Seedance 2.5 lists timestamp-level audio editing. On Sume you can detach a clip's audio and rebuild it from parts with a replaced line in the middle.
- Swap a product's colorway in a finished clip: Gemini Omni edit
One approved ad, five colorways: gemini-omni-flash-1.1 on Sume edits a source video from a prompt. Input rules, what it cannot carry, and a batch plan.
- SynthID not detected in Gemini: that does not mean the video is real
Gemini's checker only recognizes Google AI content, so no watermark means no Google mark was found, not that a person filmed it. Know its limits.
- Talking-head ad: Omni audio, Kling audio on, or Avatar Video?
Pick by length and face. Avatar Video runs 4 to 60 s from a script; Omni makes 3 to 10 s with native audio; Kling 3 adds audio at $0.21 a second.
Written by Sume