Swap a character in video: Decart Lucy reference vs Sume Recast

To swap a character in a video you have two shapes of API. Decart's Lucy models take a source video plus a text prompt, a reference image, or both, and its overview says they can swap characters. Sume's H3 Max Recast is narrower: one source video of 5 to 30 seconds and 1 to 4 reference photos, one photo per new person, with the prompt optional.

Decart's side comes from its Video Editing page and the Lucy Restyle 2 reference, read 2026-10-02. Sume's comes from Video Router and Video generation.

How does Decart handle a character swap?

Decart's editing page says you can drive an edit with text alone, a visual reference alone, or both together. The Restyle 2 API reference shows the multipart fields: data for the video, an optional prompt, an optional reference_image, an optional seed, a 720p resolution, and enhance_prompt (default true). The pages read for this post do not say how many people a reference can swap or how identity is kept across a long clip, so this post claims neither.

How does Sume's Recast work?

Recast replaces the main people in the source video with the people in your photos, left to right by default, and keeps the source length, motion, camera, cuts, and soundtrack. You pass video_url and reference_image_urls on the Video Router, or one video_url plus 1 to 4 image_url entries as input_references on /v1/videos. The resolution defaults to 768p, and the duration you send has to be the inspected source length rounded up.

A prompt can say who becomes whom or what else to keep, but the photos carry the identity.

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/person-a.jpg"],
    "resolution": "768p",
    "duration": 12,
    "mode": "async"
  }'

Which fits a multi-person scene?

If the clip has two or three people and you have a photo for each, Recast maps them directly, which is its design. A single optional reference image, as on the Decart pages read here, leaves the multi-person mapping to the prompt, and the pages do not describe it. For a long single-subject clip, Decart's unlimited duration is the clear advantage; Recast stops at 30 seconds.

Neither route is a face-swap in the legal or product sense. Sume's face-swap beta applies a ready avatar face and is a separate endpoint.

What should you test first?

Run your hardest shot: a profile turn or a fast cut. Compare identity, motion, and the audio on both. The vendors publish no quality number for this, and neither does Sume, so your own clip is the only evidence that counts.

What does Recast refuse?

Recast does not accept aspect_ratio, frame_images, generate_audio, bitrate_mode, seed, audio references, or model_params, and the validator names the field when you send one. Resolution must be 768p or 1080p, and duration has to match the inspected source length. These are not suggestions: they are checks that run before the job is created, so you will see them as request errors.

A tighter contract is a trade. You give up free-form control, and in return a bad request fails fast and the price is predictable per second by resolution: the catalog lists $0.30 per second at 768p and $0.45 at 1080p as the provider list dated 2026-10-01, with Sume billing list times 1.25. Decart's pages read for this post publish a per-second price only for the editing models, so the swap itself has no Decart number to put beside it.