Black Forest Labs has released FLUX Video Edit [fast] — a video-to-video API that makes targeted edits to a finished clip based on a text prompt without regenerating the scene. The model only changes what is specified in the prompt — objects, text, dialogue, background — while keeping the rest of the frame unchanged. The new endpoint is already available.

image
image

What happened

The new endpoint /v1/flux-tools/video-edit-v1 accepts a source clip up to 15 seconds long and up to 50 MB in MP4 format (via an HTTP(S) link or in base64) along with a text prompt. In a single request, the model can add, remove, or replace objects and characters, change the background, rewrite text in the frame, alter colors, materials, and effects, replace or translate dialogue with lip-sync, and even change the event in the frame itself. Duration, framing, camera work, and audio are taken from the source clip unless the prompt explicitly changes them, and a new line of dialogue must fit within the length of the original phrase. The result does not exceed 720p at 24 fps, and rendering a 10-second clip takes about 50 seconds. The price does not depend on the type of edit: $0.03 per second of finished video, meaning a 10-second edit costs $0.30.

Context

Until now, generative video models could only change a frame along with everything in it: touching one object would cause them to rebuild faces, lighting, and composition. FLUX Video Edit targets exactly this weak point of generative video editing, where precise detail edits could not be made without affecting the entire scene. Black Forest Labs positions the [fast] version on the Pareto frontier in terms of quality-to-price ratio: according to the company's own assessment, it is second only to Wan 3 and cheaper than Omni 1.1, H3, H3 Max, and Seedance 2.5.

Why this matters for the industry

Targeted video-to-video editing with a rate fixed per second of finished video shifts the workflow from regenerating the entire scene to industrial-scale targeted editing: localization and dubbing with lip-sync, advertising variations, previsualization, and post-production edits without re-rendering. For studios and product teams, this is a step with predictable costs that can be integrated into existing pipelines. The $0.03 per second rate sets a price benchmark that competitors named in BFL's internal comparison will have to match with their pricing, and if independent tests confirm the quality, "surgical" editing will become established as a separate class of video API.

Why this matters for users

The tool can be tested right now: a single API request sends a clip and a prompt, and a finished MP4 is returned. A five-second edit costs $0.15, which allows you to immediately calculate the cost of a batch of edits for draft and intermediate materials — advertising variations, A/B tests of formats, previsualizations. A practical first step is to run tests on complex scenes with faces, camera movement, and small objects to evaluate on your own data how well the model preserves everything not specified in the prompt.

What is still unknown / limitations

Black Forest Labs' claims about the model's quality rely solely on the company's internal Bradley-Terry Elo benchmark: the Elo evaluation protocol, dataset, criteria for a correct edit, and examples of source/result pairs have not been published, so they cannot be verified externally at this time. As of publication, there are no independent comparisons with competing models, and the announcement lacks data on SLA and rate limits necessary for production integration.

Sources

Author

Look at AI, editorial team