FLUX 3 · Upscale

Video upscaling is where the cloud-vs-local argument gets spicy. Locally you've got SeedVR2 at Apache 2.0 doing genuinely impressive restoration on a 6GB card, and the RTX super-resolution nodes doing it in seconds. This node does the same job as a metered call - and it is priced like a premium service, at roughly $0.79 per output second at the time of writing. A ten-second clip is not a rounding error.

So why use it? Because "precise vs creative" is a real control, and because you might be upscaling a clip you generated on a machine that has no room to also hold an upscaler. That's the honest case. The honest verdict: check the badge before you queue, and try SeedVR2 locally first if you can.

How it works

video_url takes a ComfyUI VIDEO - anything a video loader or another PopNow video node produced. The pack re-encodes it to MP4 if it isn't already (an MP4 source is remuxed, not re-encoded, so you don't lose a generation), uploads it, then submits, polls and downloads the result. What comes back is a normal VIDEO you can save.

Billing is by the length of the source clip, rounded up to whole seconds - not by the factor. That's the detail that makes the whole node predictable: 1.5x and 3x cost the same per second, so there's no reason to upscale conservatively.

The inputs that matter

video_url is the only required media input, and the tooltip is precise about the limits: an MP4 of up to 20 seconds. Longer than that and you're splitting the clip yourself.

upscale_mode is the interesting decision. precise stays faithful to the source - it's the restoration path, closest in spirit to what SeedVR2 does. creative also rebuilds fine detail, which is a nicer way of saying it will invent some, and it costs meaningfully more per second. On clean AI-generated footage precise is usually enough; creative earns its money on soft, compressed or genuinely old material.

upscale_factor runs 1.5 to 3 in 0.1 steps, keeps the aspect ratio, and - per the tooltip - doesn't change the price. Push it to 3 if you're going to crop.

prompt is only read in creative mode, and it's optional even there. Use it to steer what gets rebuilt: "skin texture, fabric weave, keep faces unchanged" is the sort of nudge it's for. In precise mode, typing anything here does nothing, which is a small mercy.

Outputs

video for Save Video or a downstream node, and task id - the PopNow job id. Note there's no image output: if you want a still from the result, pull it from the saved file.

Installing it

ComfyUI Manager → PopNow → Install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Key from popnow.ai/api/console/keys, paste into Settings > PopNow or set POPNOW_API_KEY, restart again to pull the catalog. No downloads, no extra dependencies - the pack uses only packages ComfyUI ships with.

Where people get burned

The upload ceiling is 90 MiB for video. A long, high-bitrate source can blow past it, and the error arrives after the encode attempt rather than as a pre-flight warning. Trim first.

Then cost. At per-second pricing, a 20-second 3x upscale is the single most expensive thing you can do in this pack without touching a text-to-video model. Set Settings > PopNow > Max credits per run and treat it as a hard wall; the run is quoted first, so anything above your ceiling stops before credits are held.

And a soft-focus source upscaled is still a soft-focus source, just bigger. If the detail was never captured, creative will make something up - plausible, occasionally weird, especially around faces. That's not a bug, it's generative upscaling, and it's the same reason the local upscaling advice says to pick the model per source rather than per hype.