Realtime AI video restyle: Decart Lucy over WebRTC vs Sume jobs

If you need to restyle or edit a live video feed as it plays, Decart's realtime Lucy models do that over WebRTC; Sume does not. Sume's video endpoints are asynchronous jobs that take a finished clip and return a new one, so the closest Sume workflow is to record the stream, then run an edit or restyle on the file.

Decart's facts come from its realtime Video Restyling and Video Editing pages, read 2026-10-02. Sume's come from Video generation, which states that video generation is asynchronous.

What do Decart's realtime models do?

Decart offers two realtime models on its pages. lucy-restyle-2 transforms a live video with a text style description, and lucy-2.1 edits a live feed, with examples such as changing a shirt colour or adding, modifying, or removing objects. Both use WebRTC. Input is fixed at 25 FPS and 1280x704 (16:9). You can send a text prompt or a reference image (JPEG, PNG, or WebP) and change them mid-stream without reconnecting.

What does Sume do instead?

Sume's video generation follows a submit, poll, download flow: you submit, receive a job id and a polling URL, poll until the status is completed, then download. The Video Router's edit mode takes one video_url and a prompt and returns a new clip; Genjutsu motion transfer takes one source video and 1 to 8 reference images. Neither is documented as accepting a live stream, and the docs name no WebRTC or streaming input.

That makes Sume the right tool for post-production and the wrong one for a live overlay. If you stream, record the session, trim the part you want, and submit the file.

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: restyle-recording-001" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "Restyle as a watercolor painting. Keep the motion identical.",
    "video_url": "https://example.com/recording.mp4",
    "mode": "async"
  }'

What can neither side promise?

Decart's realtime pages do not state session duration limits, latency numbers, or pricing, so this post cannot compare cost or delay. Sume's edit mode has no published latency either: it is a queued job, and the Jobs and results page says queued is a normal accepted state. If your product depends on frame-by-frame latency, test Decart directly; if it depends on a finished, billed, retrievable file, use a Sume job.

Which should I pick?

Pick Decart for a live camera or screen feed. Pick Sume when the output is a file you will post, ship in an ad, or feed into a timeline. Many teams need both: preview live, then run the final on the recorded clip. A look that works live will not match a file edit exactly, because they are different models, so approve the final on the final path.

How do I turn a recording into a Sume job?

Save the stream to a file at a public HTTPS URL, then submit it as video_url. The edit mode defaults to 720p, takes no aspect_ratio or duration, and cannot be combined with reference images in the same request. Poll GET /v1/jobs/{id}/status until it is terminal, then read the result. Because it is a billed job, store the job id and never resubmit a paid request just because your own process timed out.

If the recording is long, plan to cut it. Sume documents source-length limits for its media tools and for Genjutsu (4 to 30 seconds), but not for the edit mode, so test the length you need before you build on it.