FLUX 3 Video Edit API: what Sume offers to edit an existing clip

Black Forest Labs documents FLUX Video Edit as part of the FLUX 3 preview, and Sume's catalog does not list a FLUX video id. The Video Router and Video generation docs read on 2026-10-03 list no FLUX entry, so check GET /v1/videos/models for a live answer. If your goal is to edit an existing clip rather than generate a new one, Sume has three documented routes: gemini-omni-flash-1.1 for prompt edits, h3-max-recast for swapping people and higgsfield-genjutsu for Motion Transfer (Video Router docs, read 2026-10-03).

This is a comparison of what each documents, not a quality ranking, and it deliberately avoids benchmarks.

What Black Forest Labs says

The FLUX 3 overview describes FLUX Video Edit as a way to modify existing clips and gives an example edit, "Remove all the pedestrians and add a flock of pigeons around the fountain", which shows structural elements being kept while the edit applies. The same page lists standard video modes at 5 to 20 seconds, video continuation at 5 to 15 seconds, 24 fps and, at HD, $0.17 per second for text or image to video and $0.41 per second for continuation. It also calls FLUX 3 a preview model (read 2026-10-03). The page's pricing section as read states no edit price, so none is quoted here.

Preview status matters for planning. Features, limits and prices on a preview can change, and the page is the only authority.

What Sume documents for editing a clip

Sume's three source-editing routes split the job by what you want to change. Each takes a source video_url, and each is an explicit pick: sume/auto routes edits to Gemini Omni Flash 1.1 and never to Recast or Genjutsu.

Doing the pedestrian example on Sume

The closest Sume equivalent to a prompt edit like removing or adding things in a scene is Omni edit, whose own documented example is "Replace the bottle with an apple. Keep everything else the same." The request below follows that shape. Output follows the source clip and the family is 3 to 10 seconds, so a 20-second plaza shot would need to be cut first.

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: omni-plaza-001" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "Remove the pedestrians and add pigeons around the fountain. Keep everything else the same.",
    "video_url": "https://example.com/plaza.mp4",
    "resolution": "720p",
    "mode": "async"
  }'

A fair way to test two routes

If you do have access to a FLUX preview and to Sume, test with a protocol rather than an impression. Pick one 8-second clip with a clear change in mind and write the instruction once. Run it through each route you can reach, at the lowest resolution that shows the result, and keep the settings. Judge three things: did the edit happen, did everything else stay the same, and did the audio survive. Those are the three failure modes of clip editing, and they are independent, so a route can win one and lose another.

On Sume, keep the prompt style the docs use: describe the change and then say what to keep, as in "Keep everything else the same." Run Omni edit once at 720p, where the list is $0.10 per second and about $0.125 through Sume, so an 8-second test is about $1.00. If the task was a person change, run Recast at 768p as well; an 8-second test is $3.00.

Choosing without a benchmark

Pick by capability and constraint. If you need an edit longer than 10 seconds, FLUX's page and Sume's Omni family differ and you should read both. If the edit is who appears in the clip, use Recast and its photo inputs, which Omni cannot take. If you have a hard requirement for a single vendor's model that Sume does not list, the honest answer is that you will call that vendor directly for it.

Do not infer quality from this table. Run the same short clip through each route you can access, look at the results and decide, and record the settings so you can reproduce the choice.

• Check the live catalog before you design around a model.
• Treat any preview feature as subject to change.
• Compare on your own footage, not on a demo reel.