FlashVSR Video Upscale

Video upscaling is the "more pixels over time" problem, and it's the one with the least settled answer in the local ecosystem: every frame has to agree with its neighbors, so you can't just tile a still-image upscaler across the clip and hope. FlashVSR is WaveSpeed's answer - a fast video super-resolution model that runs in the cloud. This node is the whole thing: feed it a video, pick a target resolution, get a sharper video back. That's it. No model files, no VRAM, no per-frame tiling juggling.

The appeal is obvious when you've done this locally. Upscaling a video means either a heavy generative model grinding through hundreds of frames or a clever two-pass pipeline with GPU constraints and half an hour of setup. This node moves all of that off your machine and bills you per second of output instead. If you're on a laptop, or your footage is longer than your VRAM, that trade is very attractive.

The inputs (there are two)

• video - your source clip. It must be MP4 (the node validates this) and between 5 seconds and 10 minutes. Nothing exotic: a generated clip from another video node, a screen recording, old footage - all fine.
• target_resolution - 720p, 1080p, 2K, or 4K. The price badge scales with the target, because a 4K pass is genuinely more work than a 1080p one.

How it works

ComfyUI uploads your video to WaveSpeed's API via the Comfy proxy, creates a FlashVSR task, and polls the result endpoint until the enhanced video is ready, then downloads it. Needs a Comfy account with credits and an internet connection. The upload is the part people underestimate - a long 1080p clip is a big file, and it has to leave your machine before any of the fun starts. Keep clips under a few minutes unless you're patient.

Where it fits, honestly

The community's upscaling playbook is worth quoting here, because it places FlashVSR precisely: SeedVR2 beats FlashVSR on heavy upscales - the "256px to 1024px, rebuild the detail from nothing" job - and FlashVSR is the cheaper, faster option for footage that's already decent and just needs to be bigger and cleaner. The distinction is the same one that runs through all upscaling: more pixels versus more detail. FlashVSR is a recovery model, not a hallucination engine. On a soft, compressed, already-watchable clip it's excellent value; on a 240p potato it will do what it can, but the model that invents eyelashes where there were none is SeedVR2, and that one runs locally.

The sibling node, WaveSpeed Image Upscale (SeedVR2/Ultimate), covers the still-image side if you need both. For the video case, the one-liner verdict: if your footage is watchable and you want it big, this is the smoothest path from a ComfyUI graph to a 4K file on disk. It shipped in core in early 2026, so it's new enough that the exact pricing will move - check the badge on the node, not the docs.