AI Video Super-Resolution
Low-res footage, regenerated into 4K after the fact.
On August 21, 2026, Black Forest Labs (BFL) released FLUX Video Upscale, a new endpoint that enhances existing video. Instead of simply interpolating pixels, it uses BFL's latest model, FLUX 3, to "regenerate" the footage, taking sources of 480p or higher up to 4K. It adds a new option for both cleaning up rough generated video after the fact and rescuing old low-resolution footage.
Why video upscaling, why now
Two things converge here: FLUX 3 Video's own output ceiling, and the weakness of traditional upscalers.
On July 23, 2026, BFL announced "FLUX 3," a single model spanning image, video, audio and robot-action prediction. It has 12.4B parameters, generates up to 20 seconds of video with native audio, and BFL claims it was preferred over Runway Gen-4.5 in 77% of human comparisons on 720p, 10-second clips (a vendor-run, non-public evaluation that has not been independently reproduced).
Yet even once FLUX 3 Video reached general availability in early August, its native output topped out at 1080p. 2K, 4K, and open weights stayed marked "coming soon." For anyone trying to repurpose low-resolution shot or generated footage for large screens or distribution, that ceiling was a real constraint on what the model could be used for.
| Traditional interpolation upscalers | FLUX Video Upscale |
|---|---|
| Just guesses what sits between existing pixels | FLUX 3 regenerates the higher-resolution pixels |
| Blur gets more visible the higher you scale | Rebuilds smudged faces and gridded noise on water/grass |
| Source roughness is simply enlarged | Handles sources from 480p up, output up to 4K |
| One-size-fits-all processing | Two modes, precise / creative, for different needs |
A regeneration approach, with two modes to choose from
Because it regenerates rather than interpolates, you can trade fidelity for creativity by mode.
The endpoint is POST /v1/flux-tools/video-upscale-v1. You submit the video as an HTTP(S) URL or base64, then poll the returned URL until it's ready and download a signed mp4 — an asynchronous submit-and-poll pattern. There are two modes: precise (creativity: 0, keeps the source faithful and sharpens it — for material where identity matters, like faces, products, and real footage) and creative (creativity: 1, the default — restores and invents fine detail more aggressively, giving the sharpest results on generated footage, textures, and crowd scenes). The higher the source resolution, the smaller the factor needed to reach 4K — a 2K source gets there at 1.5×. Pricing is billed per output megapixel-second (megapixels per output frame multiplied by output duration in seconds), not by input resolution, factor, or whether a prompt was used. Rejected clips aren't charged.
Submit
Send the video as a URL or base64, up to 20 seconds. Specify a mode (precise / creative) and an upscale factor (1.5–3×).
Regenerate
FLUX 3 rebuilds the higher-resolution pixels from scratch. Unlike plain interpolation, it also fixes smudged faces and gridded texture noise along the way.
Retrieve
Once the status turns Ready, download the mp4 from the signed URL. Billing is based only on output megapixel-seconds.
What it means for designers, and for marketers
For both, the shift is the same: resolution becomes something you can decide on later.
Re-deliver older assets
Brand video and archival footage shot or delivered at 1080p or below can be reworked into a 4K deliverable in precise mode, without a reshoot, while keeping likeness and brand identity intact — ready for large-format displays and other high-resolution uses.
Revive low-res creative
UGC or old ad footage that sat unused for lack of resolution can be rebuilt, texture and all, into high-resolution ad creative using creative mode — growing the usable asset pool for less than reshooting or regenerating from scratch.
Clean up generated video
The gridded noise on water or grass and the smudged faces typical of text- or image-generated video can be fixed as a finishing step, without changing the generation model itself — one more way to cover for a generator's rough edges.
Resolution is no longer fixed at the moment of capture.
It has become a variable you can choose later.
The near-term outlook, and what to do first
FLUX 3 Video's own "native 2K/4K generation" is still marked coming soon, so for now generation and upscaling live on separate endpoints. Whether — and how — that gap closes, and what happens to pricing once it does, is worth watching for in BFL's next announcements.
Pilot precise mode on one asset
Run a single flagship asset through precise mode first, and check that logos and faces stay recognizable before building it into a workflow.
Work backward from delivery resolution
4K output is billed by output megapixel-second and costs noticeably more than 1080p. Check the actual resolution you need to deliver before choosing an output size and mode.
Set a rule for switching modes
For cleaning up generated video, default to creative mode, but switch to precise for shots involving faces or products — and decide that rule ahead of time.
What comes with a regeneration-based approach
Regeneration-based upscaling is also, by definition, a process that invents pixels that never existed. Creative mode aggressively "invents" texture and detail, which risks adding unintended detail to material where accuracy matters — brand logos, or the face of a real person. That BFL ships a separate precise mode looks like a direct answer to that risk.
Cost scales with output resolution and duration through the megapixel-second pricing model, so 4K and longer clips get noticeably more expensive. The 20-second cap per request also means it isn't suited to upscaling long-form content in one pass.
The 77%-preferred-over-Runway-Gen-4.5 figure cited above is BFL's own claim, from a vendor-run human comparison on 720p, 10-second clips, with no independent third-party verification confirmed. The same caveat applies to FLUX Video Upscale's own output quality: no independent third-party benchmark has surfaced yet, so this is still a case of testing it in real workflows before drawing conclusions.