Black Forest Labs' FLUX Video Edit Surgically Changes Clips for $0.03 a Second

Black Forest Labs has added a video editing endpoint to its FLUX 3 Video lineup, and the pitch is unusually narrow: send a clip, describe what should be different, and get the same clip back with only that change applied. FLUX Video Edit [fast] is a video-to-video model rather than a generator, so length, framing, camera movement, timing and audio all come from your source footage instead of being resynthesized from scratch.

The company claims it is the fastest and cheapest video editor on the market, and in their internal Bradley-Terry Elo benchmark it sits on the Pareto frontier for quality versus cost, trailing only Wan 3 while undercutting Omni 1.1, H3, H3 Max and Seedance 2.5 on price.

What the model actually does

The docs lay out eight edit categories that all fit into a single natural-language prompt:

• Remove, add or replace objects and characters, with the fill matched to the surrounding scene and stable across frames
• Rebuild the setting by swapping backgrounds or entire locations while keeping the subject and action in place
• Rewrite on-screen text like signage, stencils and lower-thirds, preserving type, angle and perspective
• Change colors, materials, lighting and weather effects (snow that actually settles on surfaces rather than falling in front of them)
• Change or translate dialogue with lip sync, with mouth movement following the new words
• Restyle the whole clip, turning live footage into watercolor or a cartoon into photoreal footage
• Alter the event in the clip, like a gull swooping in to steal a fish
• Stack several edits into one prompt or chain results as new sources for staged changes

Black Forest Labs argues that competing models often drift, altering parts of the original video even when the prompt asks for a specific edit. Their positioning is that FLUX Video Edit leaves everything unnamed exactly as shot. Faces, products, and background details all stay put unless you call them out.

Pricing, speed and limits

The economics are straightforward: $0.03 per second of output video, and the output always matches the source length, so a ten-second edit is thirty cents no matter how ambitious the prompt. A ten-second clip renders in roughly 50 seconds. Charges are output-only, meaning input resolution, prompt length and audio passthrough are not billed separately, and rejected clips cost nothing.

The hard caps are worth knowing before you wire it into a pipeline:

• Sources up to 15 seconds and 50 MiB, sent as an HTTP(S) URL or base64-encoded MP4
• Each side at least 160 pixels, and at least 17 frames after normalization to 24 fps
• Output is capped at 720p; larger sources are downscaled (a 1920x1088 source came back at 1248x704 in their example)
• Duration, resolution and aspect ratio follow the source and cannot be overridden; passing mode, seed, duration, resolution, aspect_ratio or generate_audio returns HTTP 422
• New dialogue must fit the length of the line it replaces, or the model will not render all the words

Using a separate video, image or audio clip as a reference is not supported yet but is on the roadmap.

Calling the API

The API follows a submit-and-poll pattern with only two required fields, video and prompt:

curl -X POST "https://api.bfl.ai/v1/flux-tools/video-edit-v1" \
  -H "Content-Type: application/json" \
  -H "x-key: $BFL_API_KEY" \
  -d '{
    "video": "https://example.com/harbor.mp4",
    "prompt": "Remove the orange bucket."
  }'

The submit call returns an id and a polling_url. When the status flips from Pending to Ready, result.sample holds a signed MP4 URL, which expires about an hour after generation. The optional safety_tolerance flag ranges from 0 to 4, defaulting to 2. Source audio is carried through untouched unless the prompt asks for a dialogue or sound change.

Where it fits

The obvious use cases are the ones that used to require a reshoot or a compositing seat: swapping a product or colorway across a finished ad, localizing dialogue with lip sync from a single master, removing a stray crew member or logo from a locked shot, or restyling creator footage into on-brand variants without touching the original performance. Because the output stays capped at 720p, anything destined for broadcast will need a pass through FLUX Video Upscale afterward.

The narrower framing is also the interesting one. Most of the video generation race has been about making longer, prettier clips from scratch. A model that refuses to touch anything you did not name is a different bet: that the more valuable primitive for real production work is surgical editing on footage you already shot, rather than another text-to-video sampler.