zhenzhen-Topaz 视频高清修复（平价小屋）

There's a fork in the video-upscaling road, and picking the wrong branch wastes more time than picking the wrong settings. One branch generates detail that wasn't there - SeedVR2, FlashVSR, and the whole local restorer family. The other branch restores what was there and got mangled: cleans compression noise, sharpens edges, resolves real texture one or two steps up.

This node is the second branch. It calls Topaz, the paid commercial benchmark that open restorers keep getting compared against, and it's the pack's route to it without a subscription sitting next to your ComfyUI install.

What it does

You hand it a clip, it hands back a clip. The node posts the URL to the relay, polls, downloads the result, and returns a ComfyUI VIDEO - the pack's video type, which connects to the usual save/combine nodes. video_url, task_id and response come out alongside it; task_id is your lookup key on the provider's site if the download half-fails.

There is deliberately no prompt, no duration, no frame rate in the request. This isn't generation, and the payload reflects that.

The inputs

• video_url - a public HTTP(S) MP4 link, e.g. a result from an earlier generation node. The validator checks that it actually parses.
• input_video - the same thing as a local VIDEO from your graph. You supply exactly one of the two; both or neither is an error and the node says so.

And the two that decide your result:

• resolution - 720p, 1080p, 2K, 4K. Per the tooltip, this is an output pixel budget that preserves the source aspect ratio. A vertical clip stays vertical; you're choosing how much money to spend on pixels, not a canvas.
• model - ignore the label, read the tooltip: it's the restoration model, submitted upstream as metadata.quality. The options are Ultra, Max, High, Medium, Low. Start at Max; go to Ultra when the source is genuinely rough and you've made peace with what aggressive restoration does to grain and skin texture. Drop to High or below when the source is already clean and you just want it bigger - you'll save money and lose nothing.

api_config takes the key from a zhenzhen-seedance2-low-price-api-settings node, or the SEEDANCE_API_KEY environment variable. seed is a ComfyUI cache control and, per its tooltip, is not sent to Topaz at all - the practical use is leaving it fixed so re-running with an unchanged graph doesn't re-bill the same clip while you experiment elsewhere.

Install

Manager → search zhenzhen, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen
# restart ComfyUI

Nothing local to load - no model files, no VRAM. The first start installs aiohttp, transformers, opencv-python and the rest of requirements.txt, and the pack starts a helper web app on port 8080 to serve its front-end assets. You'll need a key from the provider's site before anything runs.

Where people get burned

A create request that times out may still be running, and the node will not resubmit it. That's intentional - replaying the submission would risk two charges for one clip - and the error message sends you to the provider's task list to check before retrying. Read it, check, then decide.

Don't feed it mush. Restoration has a floor. A 256p source is not going to come back as crisp 1080p; you'll get confident edges around genuinely absent detail, which reads as a waxy oil painting on skin. That case is the other branch: use a generative restorer, and if you need one in this same pack, the FlashVSR upscale node covers the cheap 480p-source case the author built it for.

Face drift is the medium, not the node. Any restore pass alters faces slightly, and on a long clip that's where you notice it. If identity matters more than texture, budget for it - upscale the plate, not the performance.

Interlaced or odd-frame-rate sources. Restoration tools assume progressive frames. Drop old camcorder footage in and you'll get combing that no quality tier fixes; deinterlace first.

And if you have a folder of clips rather than one: this node waits for each result in turn. The concurrent submit twin in the same pack pushes ten into a shared pool at once.