An AI Video Upscaler for 4K: FLUX Video Upscale

How this AI video upscaler handles each frame

FLUX Video Upscale takes one clip and returns a larger one. What happens in between is not stretching: the model regenerates each frame at the target size.

Super-resolution rebuilds detail, not stretched pixels

Super-resolution means the model rebuilds the frame at the new resolution. A plain scaler stretches the pixels that already exist, so the result looks soft or blocky. FLUX Video Upscale builds detail at the target size rather than reusing those pixels.

The factor can be set anywhere from 1.5x to 3x, with 2x as the default, and the output keeps the source aspect ratio. In creative mode the detail is invented, so the output can contain texture the source never held. The mode setting controls how far that goes.

It repairs smudged faces and grid artifacts on textures

Enlarging a clip makes flaws visible that were hidden at the original size. Black Forest Labs names three: smudged faces, soft eyes, and the gridded artifacts that appear on textures such as water and grass.

The model is tuned for the output of FLUX 3 Video: 20-Second Clips With Native Audio, and the same component already runs as the 1080p step in that generation pipeline. It also takes footage that FLUX 3 Video did not generate, within the source limits further down, so existing clips can go through it.

Precise or Creative: which mode fits your footage

One integer, the creativity parameter, selects the mode. Footage with people or products in frame should use precise. Generated footage, textures, crowds and scenery suit creative, which is the default.

Precise keeps faces and products intact

Precise preserves the source exactly and sharpens it. It runs 4 steps at the lower rate of $0.07 per megapixel-second, and Black Forest Labs recommends it when identity or reference details must stay consistent. Footage of products, brand assets or real people belongs in this mode.

Creative is the default, and it can alter details

Creative runs 8 steps at $0.10 per megapixel-second and restores and invents fine detail more aggressively. It is the mode applied when the request leaves the parameter out.

Black Forest Labs states that creative can change or replace identity, and that reference consistency is lower than with precise. Faces and products can drift. A clip with a face or a product in it should therefore be sent with creativity set to 0.

What upscaling video to 4K costs and what it won't accept

Two things matter before you submit a job: the rate per megapixel-second, and the source limits that decide whether it runs at all. The prices below are as of August 23, 2026, from the Black Forest Labs docs.

Pricing is per megapixel-second of output video

A megapixel-second is the megapixels per output frame multiplied by seconds of output. One megapixel is 1,048,576 pixels (1024 x 1024), the same definition Black Forest Labs uses for FLUX image pricing. Only the delivered file enters the calculation.

A 10-second clip delivered at 1080p has frames of 1.98 megapixels (1920 x 1080), which gives 19.8 megapixel-seconds. That bills at $1.39 in precise mode or $1.98 in creative mode.

The source resolution, the upscale factor and any optional prompt carry no charge of their own, and rejected clips are not charged. A cost estimate therefore needs only the mode, the output resolution and the duration. If the response you get back is JSON, formatting it makes it easier to read.

Sources over 20 seconds or 2560x1440 are rejected

Source clips are limited to 20 seconds, 50MB and 2560x1440 (2K). The docs state that sources longer than 20 seconds or larger than 2560x1440 are rejected before processing. They are not truncated or downscaled, and no charge applies.

The 2560x1440 limit means the endpoint upscales toward 4K and does not take 4K in. A full 2K source reaches 4K at the minimum 1.5x factor.

Output frames are capped at about 14.4 megapixels, so a very large source asked for 3x is upscaled by less than the requested factor. The output keeps the source aspect ratio and the original audio track.

Where upscaling fits in your video pipeline

The meter runs on delivered output, so where upscaling sits in the edit decides how many seconds you pay for. Black Forest Labs gives two tips, and the 20-second limit adds a third step. If you are still generating the footage, FLUX 3 Video: 20-Second Clips With Native Audio and Seedance 2.5 by ByteDance: 30 Seconds in One Pass cover that step.

Upscale last, after the cut is locked

Run the upscale after editing and trimming, so the meter only counts footage that stays in the cut. Upscaling first means paying for frames the edit later removes.

Start from the least compressed source available. Compression artifacts limit how much real detail the upscaler can recover. A source longer than 20 seconds has to be split before upload, because it is rejected.

Download the result within an hour

Submitting a clip is one POST followed by polling, and the result is fetched from the polling URL. The signed delivery URL expires about 1 hour after the result is ready. A pipeline that queues jobs should add a download step that runs inside that window.