Black Forest Labs has released a standalone tool called FLUX Upscale that regenerates video shot at 480p or higher up to native 4K. It is not really a model upgrade — more like a single step of the company’s own pipeline being split off and sold on its own. The same approach was already tucked inside FLUX 3 Video, where it handled the 1080p output stage.
The German team built its reputation in the image world with the FLUX family of open-source models. This release sits further from image generation and closer to post-production. Anyone without API access who wants to see how regeneration differs from ordinary upscaling can drop a clip into PixPix’s Video Enhance tool for free and compare the two versions side by side — a lot more useful than parsing a spec sheet.
Regeneration is not the same as upscaling
Standard video upscaling relies on interpolation: where pixels are missing, it guesses, and the result looks smooth but empty. Black Forest Labs positions FLUX Upscale differently — as regeneration, where the model looks at the low-resolution input and redraws every frame. The flaws it targets are specific: blurry faces, water surfaces, and the grid-like artifacts that show up in high-frequency textures like grass. Those three happen to be the recurring weak points of current video generation models, which sacrifice detail during generation to save compute and then have to claw it back afterward.
The tool supports three scaling factors — 1.5x, 2x, and 3x. Starting from HD source footage, those correspond roughly to 1080p, 2K, and 4K.
Two modes, two different bills
Precise mode runs four steps at $0.07 per megapixel-second; Creative mode runs eight steps at $0.10 on the same basis. The step count doubles, the price rises by 43%, and according to the company the extra compute goes into detail reconstruction — at the cost of making bigger changes to the original image.
Rough math based on output resolution: a 4K frame runs about 8.29 megapixels. Under Precise mode, that’s roughly $0.58 per second of footage; under Creative, about $0.83. For a ten-second finished clip, the upscaling step alone costs six to eight dollars.
That number is not cheap by the standards of today’s video generation market. Over the past year, several vendors have repeatedly driven down the per-second cost of generation itself, with fractions-of-a-cent pricing now showing up in public rate cards. Upscaling, by contrast, sits on the more expensive side of the cost curve. That is not surprising: generation can front-load most of its compute-heavy work at low resolution, while upscaling has to run every frame at the target resolution — pixel count is a hard constraint there is no way around.
Billing by pixel-second carries another consequence: cost scales linearly with duration, with no economies of scale. A one-minute clip costs six times as much as a ten-second one.
Who ends up paying for this
The visible use cases cluster in a few places. Advertising and e-commerce footage is typically short — a few seconds to a bit over ten — so per-clip cost stays manageable while image quality still matters. Trailer clips for film and short-form drama fall into the same bucket. Then there is legacy footage restoration: teams sitting on a backlog of older clips find it cheaper to regenerate than to reshoot.
On the e-commerce side, tools have already turned this step into a button. PixPix puts Video Enhance and image Upscale on the same surface as product photo sets and video cloning — a blurry hero shot or a shaky product video gets fixed on the spot, and the free tier is enough to process a full product video. For small and mid-size sellers, a pay-per-pixel-second API is really built for teams already running at volume; trying the effect on a tool like this first, then deciding whether precision is worth paying for, is the more practical path.
What Black Forest Labs did not disclose this time is also worth noting: no processing latency figures, no stated duration cap per job, and nothing about how it performs on video that was not generated by FLUX in the first place. The tool is already live in the official playground and API docs — how it actually performs will need a batch of people running real footage through it.
For video teams working in China, the bigger takeaway here is that it confirms a direction: competition among video models is shifting from whether they can generate footage at all to what happens after generation — the extra passes still needed. Resolution, duration, and consistency, once bundled into a single model metric, are starting to split off into services that get priced separately.
Sources: Black Forest Labs official blog, CocoLoop, FLUX tool documentation and official pricing page; per-megapixel-second pricing, step counts, and the three scaling factors for both modes are verified against official pricing.