FLUX Video Upscale

Most of the AI video you generate comes out soft - fine for a first pass, embarrassing for anything you'd show someone. FLUX Video Upscale is Black Forest Labs' hosted fix: feed it a video and it restores and sharpens detail frame by frame, in the cloud, from inside ComfyUI. If you've been rendering at 720p and upscaling in a video editor, this node is the version that lets the model decide what detail to invent instead of a dumb lanczos filter guessing.

The mechanism is generative upscaling, which behaves differently from the classic image upscalers the KB covers. A traditional upscaler interpolates pixels; a generative one looks at each frame and reconstructs plausible high-frequency detail, which is why BFL gives you a creativity control instead of a sharpness slider - a lower value mostly restores what's there, a higher value lets the model invent fine texture. The upscaling doc's lesson applies hard here: generative upscaling can introduce or invent details that weren't in the source, so check faces and fine geometry rather than assuming the output is truth. positivePrompt is optional but useful for steering what gets enhanced - describe the subject and the enhancement follows it.

The inputs are short because this is a single-purpose node. video is required, fed as a UUID or URL of the source clip - in practice you'll get that from another Runware video node's output or an uploaded clip. upscaleFactor defaults to 2 (doubling width and height), and the tooltip carries the honest cap: output frames top out around 14.4 megapixels, so a very high factor on a big source just runs into the ceiling. settings.creativity defaults to 1. There's a settings.safetyTolerance (default 2) and the usual outputFormat/outputQuality group. The one thing to notice is the output type: this returns image, not video - it's grouped under Runware's image tasks, and the upscaled result lands as frames you'd feed back into a video-assembly or save step rather than coming back as a playable VIDEO.

Install

The shared Runware pack setup applies: search Runware in ComfyUI Manager and install, or clone https://github.com/Runware/ComfyUI-Runware into custom_nodes plus pip install -r requirements.txt, then restart. Add your Runware API key (Settings → Runware API key, RUNWARE_API_KEY, or runware auth login) - no checkpoints involved.

When it earns its keep

Reach for it when your final clip needs to survive a bigger screen and the source model couldn't render native 4K - that's the case where generative enhancement genuinely beats re-rendering at a higher resolution, which would cost more and change the shot. It's per-call billed on video, so run it last, on the clips you're keeping, and keep creativity modest unless you explicitly want the model to dream up texture. Your source video leaves the machine and passes through Runware's moderation; if the footage is sensitive, that's the dealbreaker to weigh.