Black Forest Labs has released FLUX Upscale, a standalone API that rebuilds video at a higher resolution. Footage from 480p upward can be lifted by a factor of 1.5x to 3x, reaching native 4K at the top end. Rather than interpolating between existing pixels, the endpoint regenerates each frame with FLUX 3, and a single parameter decides whether you get fidelity to the source or aggressive detail work.

Regenerating rather than stretching

A shot delivered in HD is not always ready to ship. Broadcast, campaign work, and big-screen playback all demand more pixels than that, and upscalers running on local hardware tend to lose quality once they reach those resolutions. The high-quality alternatives, meanwhile, are slow when you push video through them at scale. That is the gap the company describes as its starting point.

FLUX Upscale raises resolution by regenerating the footage with FLUX 3. Because it is not a simple enlargement, it can repair the flaws that show up in generated video, such as smudged faces or gridded artifacts across textures like water and grass. The model understands FLUX 3 output natively, but it accepts ordinary footage that was never generated by AI.

Precise for fidelity, Creative for detail

Behavior is controlled by the creativity parameter. Setting it to 0 selects Precise mode, which runs 4 steps and sharpens the footage while preserving the source exactly. That is the choice when the subject must not change: faces, products, brand assets, and footage of real people.

Setting it to 1 selects Creative mode, which runs 8 steps. It restores detail more aggressively and invents what is missing, which works well on generated footage, textures, crowds, and scenery. Identity is not strictly preserved, so faces and products can drift. A short prompt describing the clip steers Creative mode toward the kind of detail you want.

Input caps out at 20 seconds and 2K, output at about 14.4 megapixels

Input arrives either as a base64-encoded mp4 of up to 50MB or as an HTTP(S) URL. Clips are limited to 20 seconds and 2560x1440, and anything above those limits is rejected outright rather than trimmed or downscaled. Rejected clips are not charged. The upscale_factor parameter accepts values between 1.5 and 3 and defaults to 2. Aspect ratio is preserved, and the source audio track survives the process.

There is a ceiling on the output side as well: roughly 14.4 megapixels per frame. Request 3x on a source that is already high resolution and the real factor lands below what you asked for. The minimum 1.5x exists to carry a full 2K source up to 4K, which is to say the endpoint upscales toward 4K rather than accepting 4K as input.

Calling it is one POST plus polling until the result is ready.

curl -sS -X POST "https://api.bfl.ai/v1/flux-tools/video-upscale-v1" \
  -H "Content-Type: application/json" \
  -H "x-key: $BFL_API_KEY" \
  -d '{"input_video": "https://example.com/source-clip.mp4", "upscale_factor": 2.0, "creativity": 1}'

Poll the polling_url that comes back, and once the status flips to Ready you receive a signed mp4 URL. That delivery URL expires after about 1 hour, so any pipeline built on this needs to download the result immediately rather than queue jobs and collect them later.

Pricing is output area multiplied by duration

Billing is measured in megapixel-seconds: megapixels per output frame times the length of the output in seconds. Precise mode costs 0.07 USD (about 11 yen) per megapixel-second and Creative mode costs 0.10 USD (about 16 yen). You are charged only for output that is actually delivered, and one megapixel is defined as 1,048,576 pixels, matching FLUX image pricing.

Converted to a per-second figure by output resolution, 1080p runs 0.14 USD (about 22 yen) in Precise and 0.20 USD (about 32 yen) in Creative. At 2K the two modes cost 0.25 USD (about 40 yen) and 0.35 USD (about 56 yen), and at 4K they reach 0.55 USD (about 87 yen) and 0.79 USD (about 126 yen). A 10-second clip taken to 1080p counts as 19.8 megapixel-seconds, so 1.39 USD (about 220 yen) in Precise and 1.98 USD (about 315 yen) in Creative.

※1 USD = 159 JPY (as of August 21, 2026)

Running the upscale as the final step, after editing and trimming, avoids waste. Upscale footage you later cut and you have paid for frames nobody will see.

Summary

FLUX Upscale is a new Black Forest Labs API that regenerates footage from 480p upward at up to 4K. Two modes, Precise for fidelity and Creative for detail, are selected through the creativity value. Input is limited to 20 seconds, 50MB, and 2K, output is capped at roughly 14.4 megapixels per frame, and pricing follows the output area multiplied by duration. The less compressed your source, the more real detail there is to recover, which makes this worth planning for as early as the export stage.