VOSR2 2K 视频超分

Video upscaling is the expensive corner of the upscaling world: per-frame image upscalers fall apart on motion, and the models that keep consecutive frames consistent are the ones that eat your whole GPU. So this node's whole pitch is "let someone else's GPU do it." VOSR2_Video_Upscale (display name "VOSR2 2K 视频超分") is an API node inside ComfyUI_Seedance: you connect a local video or paste a public URL, and the api.seedance.nz gateway runs a vosr2-video-upscale model server-side and returns a 2K video into your graph.

Why you'd reach for it

If you already run SeedVR2 or FlashVSR locally for video, this is probably not an upgrade - it's a paid alternative with a fixed output. Where it earns its keep is the situation the community actually hitches on: you have footage and no card big enough to upscale it in reasonable time, or you're already inside this pack's ecosystem for its cloud video generators and adding one more API model costs you nothing new to learn. You get a clean VIDEO output, a progress bar, and none of the temporal-consistency headaches that make local video SR fiddly - that part is the upstream model's problem, not yours.

The honest caveat, same as every API node: your source video is uploaded to a third-party server, and each job is metered. For a single clip you don't care; for a hundred, do the math first. And there's no quality/size control here - the output is 2K, period.

How it works

You give it exactly one source. Either connect a ComfyUI VIDEO to input_video, which the node uploads automatically (it handles mp4, mov, avi and mkv containers), or type a public video URL into video_url. Then it submits POST /v1/video/generations with model=vosr2-video-upscale and the source tucked into metadata.video_url, polls the task, and downloads the 2K result as a standard VIDEO. Set both inputs and it errors - one source only. Set neither and it errors too, before you've paid for anything.

The inputs and outputs that matter

• input_video - a local ComfyUI video. The most common choice; drag in a Load Video output.
• video_url - a public, direct-link video URL. Use this when the clip is already hosted (S3, Imgur-hosted mp4, whatever) and you don't want to upload. Leave it empty when you connect input_video.
• api_config - optional; connect a Seedance API Config node, or let it use SEEDANCE_API_KEY / config/.env.
• skip_error - when on, failures return a placeholder video instead of aborting the workflow.

Outputs: video (the 2K result - feed it to SaveVideo or a preview node), plus video_url (the temporary result link), task_id, and response for logging.

Installing it

The node ships inside ComfyUI_Seedance, so install the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/ComfyUI_Seedance.git
cd ../.. && python -m pip install -r custom_nodes/ComfyUI_Seedance/requirements.txt

Restart after that. In ComfyUI Manager, search "ComfyUI Seedance" or seedance and install from there - same result. No model files, and the only Python dependency is requests. You need an API token from api.seedance.nz/console.

Common issues

• "VOSR2 accepts exactly one source" - you filled video_url and connected input_video. Pick one.
• "connect input_video or provide video_url" - both were empty. The URL field defaults to blank, so a forgotten connection is the usual cause.
• The public URL didn't work - it must be an http(s) direct link, not a YouTube or Drive page. The node validates the scheme up front.
• Missing API key - connect Seedance API Config or set SEEDANCE_API_KEY before launching.
• Windows SSL errors - a known pack quirk; update requests/certifi, or set SEEDANCE_CA_BUNDLE. There's also SEEDANCE_SSL_VERIFY=0, but treat that as a temporary debug lever.
• Result gone later - gateway links are temporary. Save the video output while it's fresh.

Upscale one clip now and again and this is a nice no-brainer. Upscale footage at scale and you'll want the concurrent sibling node instead, which fans this out to ten parallel jobs. That's the version this pack is really built around.