Swap a person in a 4-second video: Recast needs 5 seconds

A 4-second clip is one second too short for h3-max-recast, which accepts 5 to 30 seconds. You have three documented ways forward on Sume: higgsfield-genjutsu accepts 4 to 30 seconds, the Beta Avatar Face Swap targets source videos of about 4 to 15 seconds with usable audio, and gemini-omni-flash-1.1 edit works within a 3 to 10 second family (Video Router docs and Face swap docs, read 2026-10-03). Or you can re-capture a take of at least 5 seconds and use Recast, which is the only one of the four that replaces the people with photos while keeping the source's cuts and sound.

Do not pad the clip to reach 5 seconds unless the padding is something you would be happy to ship. A frozen last frame or a looped tail will be part of the footage the model re-renders.

What each short-clip route does

The four tools do different jobs, and short clips are common in social video, so it is worth knowing which job you are really doing.

Option 1: Genjutsu for a 4-second swap

Genjutsu is the closest to Recast in spirit for short clips: you send one source video and 1 to 8 reference images, and the output duration is the input video's length rounded up. It supports 480p and 720p and rejects aspect_ratio, generate_audio, bitrate_mode and audio references. Check that it appears in your catalog first, because it is listed only when its provider is configured.

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: genjutsu-short-001" \
  -d '{
    "model": "higgsfield-genjutsu",
    "video_url": "https://example.com/reaction-4s.mp4",
    "reference_image_urls": ["https://example.com/new-person.jpg"],
    "resolution": "720p",
    "mode": "async"
  }'

Option 2: face swap with a ready avatar

If the replacement is one of your own Avatar 1.0 identities, the Beta face-swap endpoint is built for short clips. It takes avatar_handle, video_url and a required quality of standard, plus or max; there is no default. It rejects prompts, aspect ratio and duration knobs, and it wants a public HTTPS source video, not a signed or private URL. The source should have usable audio, which the pipeline mixes back into the result.

The trade is control for convenience: you pick an avatar rather than an arbitrary photo, and you get a whole-person swap that keeps the source audio. If you do not already have a ready avatar, create one first, which is a separate step described in the avatar docs.

One more check before you commit to any route is the aspect ratio of the clip. Genjutsu rejects an aspect_ratio field, so the output follows the source, and face swap rejects aspect-ratio and duration knobs entirely. That is convenient for vertical social clips, since you do not need to restate the shape, but it also means you cannot fix a wrong crop in the request. Crop the source first, then swap.

Cost is the last consideration. Short clips are cheap on every route, so choose by what the clip needs rather than by a few cents: photos of any person, use Genjutsu; a ready avatar and the original voice, use face swap; a prompt-only edit, use Omni.

Option 3: re-capture longer

The cleanest fix for a 4-second Recast candidate is a longer source. If you control the footage, record 6 or 7 seconds, which also gives you handles to trim. If you do not, ask whether the missing second is worth the change of tool. A 5-second 768p Recast is $1.88 through Sume, so the price of staying with Recast is small once the footage allows it.

Whichever route you choose, probe the clip first and round the length up when you quote, since billing rounds up. A 4.01-second clip is a 5-second job on a tool that bills by the second, and a 4.99-second clip is still below Recast's floor by the catalog's rule, so do not assume rounding saves you.

• Check the catalog for Genjutsu before designing around it.
• Use face swap only with an avatar you have the right to use.
• Re-capture to 5 seconds or more if Recast is the goal.