Black Forest Labs shipped an upscaler for video: you feed it a recording from 480p up, it hands it back enlarged 1.5, 2 or 3 times, up to native 4K. Two modes, a price per megapixel-second, only through the API and the playground. No benchmarks.
- FLUX Upscale takes video from 480p up and regenerates it with a factor of 1.5x, 2x or 3x - roughly 1080p, 2K and 4K from an HD input.
- Two modes: Precise at 4 steps for 0.07 dollars per megapixel-second and Creative at 8 steps for 0.10 dollars, with more detail and less fidelity to the face.
- Access is API and playground only. There are no open weights and no published benchmarks.
I have run video through an upscaler frame by frame. On a still it looks clean. Put it in motion and the face starts to breathe, the grass crawls, the water falls apart into small squares. It flickers.
That is why this announcement took more of my time than yet another generator. Stable video upscaling is an old hole in the craft, and the few tools that plug it decently are closed and expensive.
The two modes say the most
Precise holds the face. Creative corrects it more aggressively and can hand you back a different person. That is not a small line in the documentation, it is the difference between a frame that goes into the edit and a frame that goes back for a reshoot.
The price is by size and by time at once. Working from the announced numbers, one second in 4K on Precise comes to about 0.58 dollars, so a thirty-second clip is somewhere near 17 dollars before you have tried a second version. For an advert piece that is bearable. For an archive of old material the bill quickly becomes a different conversation.
And there are no weights. The company that made its name with open models is keeping the door shut this time - you go through their API or you do not go. It is logical, the tool is aimed at production, not at your laptop. But it also means one player sets the price.
There are no benchmarks either, so the check is one and it is yours. Take your own frame with water, grass or a face in motion, run it through both modes and watch it in motion, not on a still. That is where you see whether it flickers.