FLUX Upscale gives generated video a 4K output
FLUX Upscale enlarges clips of up to 20 seconds to 4K. Two modes prioritize either source fidelity or detail reconstruction.
Video generators still often produce footage at resolutions below those required for advertising campaigns, television broadcasts, or large screens. Black Forest Labs is adding this finishing step to its API with FLUX Upscale, a tool that can enlarge a clip to 4K while correcting certain flaws commonly found in generated video.
The tool complements FLUX 3 Video, whose first public release generates clips of up to 20 seconds in HD or Full HD, with a soundtrack created at the same time. FLUX Upscale can process these results directly, but it is not limited to content produced by Black Forest Labs. Any MP4 file that meets the endpoint’s requirements can be submitted through an HTTP address or as encoded data.
The process does more than stretch each frame. Black Forest Labs says the model reconstructs the sequence at a higher resolution, targeting issues such as indistinct faces, blurred textures, and grid-like artifacts that can appear on water, grass, or repeating surfaces. However, no benchmark compares these corrections with those produced by local software or competing services.
Two modes determine how much freedom the process has. Precise mode uses four steps and aims to preserve the source’s appearance as closely as possible. It is primarily intended for faces, real people, products, visual identities, and brand assets whose details should not change.
Creative mode uses eight steps and reconstructs missing information more actively. It can enhance hair, crowds, landscapes, or hard-to-read materials, but it may also alter what it attempts to restore. Black Forest Labs explicitly warns that a person’s identity or a product’s appearance may drift. The additional detail is therefore a new interpretation of the content, not the reliable recovery of information contained in the original file.
This distinction is important for both live-action footage and generated video. Creative mode may make a slightly blurred face appear more defined while no longer depicting exactly the same person. Precise mode reduces that risk without guaranteeing that a detail absent from the source can be recovered. Increasing resolution cannot reliably reconstruct something that was never recorded.
An optional prompt lets users describe the scene’s content. This instruction can guide the type of details being reconstructed, particularly when a surface or object remains ambiguous. Leaving it blank produces a more neutral result. This option makes Creative mode closer to guided regeneration than a purely technical conversion.
Users select an upscale factor between 1.5 and 3, with 2 as the default. From a 1280 × 720 source, a factor of 1.5 produces 1920 × 1080, a factor of 2 reaches 2560 × 1440, and a factor of 3 reaches 3840 × 2160. The “2K” label used in the company’s materials remains approximate, since multiple standards use the term for different dimensions.
The endpoint documentation limits source files to 20 seconds, 50 MB, and 2560 × 1440. A video exceeding any of these limits is rejected before processing rather than being automatically trimmed or reduced. Output resolution is capped at approximately 14.4 MP per frame. If the requested factor exceeds that limit, the service applies a smaller enlargement.
A 2560 × 1440 source therefore cannot simply be multiplied by three. The minimum factor of 1.5 is already enough to bring it to 3840 × 2160. The tool is designed to reach 4K from smaller sources, but it does not accept 4K video as input for enlargement to an even higher resolution.
The aspect ratio remains unchanged. Vertical, square, and panoramic footage retains its original format instead of being cropped. The endpoint also preserves the existing audio track: it neither regenerates the audio nor resynchronizes it with the restored frames.
The result is delivered as an MP4 file through a signed address. That address expires approximately one hour after processing is completed, requiring the file to be downloaded promptly. The documentation does not yet specify the codec, bitrate, color depth, chroma subsampling, color space, or HDR support.
This lack of information currently limits any assessment of the tool within a professional finishing pipeline. A 4K output does not necessarily mean the file is immediately suitable for color grading, compositing, or broadcast delivery. Demanding workflows may still require conversion to an intermediate format, along with checks for banding, compression, and consistency across successive frames.
Pricing depends on the output resolution and duration rather than the size of the source file. Black Forest Labs charges $0.07 per MP-second in Precise mode and $0.10 per MP-second in Creative mode. The platform charges only for the file that is successfully delivered.
At the dimensions provided by the company, one second costs approximately $0.14 at 1080p in Precise mode and $0.20 in Creative mode. The price rises to about $0.25 and $0.35 around 2K, then $0.55 and $0.79 at 4K. A 10-second 4K clip therefore costs approximately $5.50 in Precise mode or $7.90 in Creative mode. At the maximum duration of 20 seconds, the cost reaches roughly $11 or $15.80.
These prices make it more practical to reserve processing for shots retained in the final edit. Black Forest Labs also recommends upscaling after cuts and adjustments have been completed, using the least compressed source available. Processing several full-length variations before making a selection could quickly increase costs.
The integration operates asynchronously. A request is sent to the `video-upscale-v1` endpoint, after which the project polls a status address until the file is ready. Webhooks can automate retrieval for services processing higher volumes. The same tool is available in the Black Forest Labs Playground without requiring an integration.
FLUX Upscale addresses a clear limitation in generative video workflows: producing a shot is not enough when its output format remains below the final project’s requirements. Its usefulness