FLUX 3 Video adds Edit [fast] mode for targeted video edits

• FLUX 3 Video Edit [fast] applies prompt-led revisions to an existing clip, covering objects, backgrounds, on-screen text, materials, effects, dialogue, and events, according to the launch post.
• BFL says the model's key behavior is edit isolation, making one requested change while retaining the rest of the shot, as the follow-up describes it.
• The endpoint accepts source clips up to 15 seconds and returns the same duration and aspect ratio at 24 fps and up to 720p, via the documentation post.
• Separate video, image, and audio reference inputs are not supported yet, though the follow-up says they are coming.
• BFL positions the fast mode as near-leading quality at the market's lowest cost in a second launch post, a first-party performance claim rather than an independent comparison.

FLUX 3 had already combined video and synchronized audio generation from text, images, and video references, as Black Forest Labs' July launch post laid out. The edit endpoint narrows that broad video-to-video idea into a short-form post-production workflow, with Runware's integration documentation listing preserved camera work, timing, framing, motion, and audio unless the prompt changes them.

The launch post divides the edit surface into six useful buckets:

1. Add, remove, or replace objects and characters.
2. Replace backgrounds.
3. Rewrite on-screen text.
4. Change color, material, or visual effects.
5. Change or translate dialogue with lip sync.
6. Change an event in the clip.

The workflow supports both a stack of changes in one prompt and iterative passes, according to the official endpoint documentation. That makes a translated line, revised end card, and colorway change part of the same source-clip workflow.

BFL says competing video models often alter unrelated areas of a source clip during a narrowly specified edit. Its stated target is a change that leaves everything else alone.

The currently unavailable reference inputs are:

• A separate video reference.
• An image reference.
• An audio reference.

Creator egeberkina singled out consistency across edits, rather than faster generation, as the feature worth testing in Higgsfield.

The API takes a source MP4 by HTTP(S) URL or base64 plus a text prompt in one POST, then returns a task ID and polling URL. Results move from Pending to Ready, and the signed delivery URL expires about an hour after completion, per the official API documentation.

Its production constraints are concrete:

• Input: up to 15 seconds and 50 MiB.
• Output: source-derived duration and aspect ratio, 24 fps, and up to 720p.
• Billing: $0.03 per output second, or $0.30 for 10 seconds at the listed rate.
• Runtime: roughly 50 seconds for a 10-second edit.

These limits belong to Edit. The wider FLUX 3 video API supports generation up to 20 seconds and full HD, with asynchronous polling across text-to-video, image-to-video, and continuation modes in BFL's video documentation.

MayorKingAI generated the same 15-second prompt through Magnific with Seedance 2.5 at 1080p, MiniMax H3 at 2K, FLUX 3 at 1080p, and Wan 3.0 Prime at 1080p.

A four-way FLUX 3, Seedance, MiniMax, and Wan comparison

It is a text-to-video comparison, not a test of targeted editing or preservation. MayorKingAI later said they preferred FLUX 3 in a reply.