Video Upscaler · Upscale

Video upscaling is the third job in the upscaling taxonomy and the one with the least settled answer. A still image is one frame you can afford to be slow on; a clip is every frame, and every frame has to agree with the ones next to it or the result crawls with flicker. Locally that's SeedVR2 on a good card, or an RTX 2-pass, and both are the sort of thing you run overnight for a thirty-second clip.

Video Upscaler · Upscale is the other route: 0.9 credits per second and up. A ten-second clip to 1080p is nine credits - under a cent per second - and it keeps the audio track, which most local video upscalers cheerfully throw away.

What it is

popnow/video-upscaler-v1/upscale, one of PopNow's own models rather than a resold frontier API, wrapped as a ComfyUI node. Video in, video out as a normal VIDEO value. No GPU work on your side, no model weights, and no dependency on any of the video node packs people fight with in the open ecosystem.

The tooltip is unusually precise about the contract and it's worth trusting it: length, motion and the audio track are kept. That's the three things that break in a bad video upscale, and the audio one is the giveaway that this is a full pipeline rather than a frame-by-frame image upscaler in a trench coat.

The inputs

video_url takes the clip, and the range is 3 to 60 seconds. Out-of-range clips are rejected before any charge, which is a nicer failure mode than most metered services manage. Clips are uploaded as MP4, and an MP4 source is remuxed rather than re-encoded.

resolution defaults to 1080p and you get four tiers: 720p, 1080p, 2K, 4K. Tier matters twice over - it's a pixel budget with your aspect ratio kept, and it sets the per-second rate. So this isn't the flat-price trick the image upscaler pulls; going to 4K genuinely costs more per second, which is the honest way to price it.

The tooltip's own worked example is the best explanation of what "tier" means here: an 832x480 clip at 1080p comes back at 1896x1094. Not 1920x1080 - the shape is preserved exactly and the tier is a budget you land near, not a target you're cropped into.

Outputs: video and task id. (No last-frame output on this one, unlike the pack's generation video nodes - if you want a still, pull one from the clip's frames afterwards.)

Install

Manager → search PopNow → Install → restart, or:

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

Then a key from popnow.ai/api/console/keys in Settings > PopNow, or:

POPNOW_API_KEY=sk-popnow-... python main.py

Zero dependencies - the pack only uses what ComfyUI already ships, PyAV included. Nothing to download, nothing to fit in VRAM. Restart again after the key is set so the node catalog is built from your account.

Where people get burned

The 90 MiB upload ceiling. Videos are re-encoded to MP4 before upload and anything over 90 MiB is refused. A 60-second 4K phone clip will blow straight through that. Trim, or transcode down before the socket - and remember that the source you upload is a re-encode, so don't feed it a pristine master and expect the exact source pixels back.

Upscaling doesn't fix a bad source. The upscaling essay makes the point that every generative rung of the ladder rewrites content. Heavy upscales of genuinely low-res sources are where these models earn their money; a clip that's already sharp but noisy mostly gets its noise restated at higher resolution. Clean first if you can.

Cost scales with both knobs you set. 30 seconds at 4K is thirty times the per-second rate of the 720p tier - check the badge on the node before you queue, because the node shows the current rate for the settings you've picked, and quotes the job before anything is held.

Cancelling isn't a refund. Interrupting a run stops ComfyUI waiting; the job keeps going on PopNow and is charged if it succeeds, with the result sitting in the API console. Set Max credits per run in Settings > PopNow if you're experimenting with long clips.