FlashVSR 480P 视频超分

FlashVSR is a real-time video super-resolution model, and normally you'd need a decent GPU and a pile of weights to run it. This node calls a hosted copy of it instead - the pack's own tests turned a 854×480, 3-second H.264 clip into a valid 1920×1024 MP4. If you've got an old SD 1.5-era render or a phone clip sitting at 480P that you want upscaled before it goes anywhere near your final workflow, this is the cheap, zero-VRAM way to do it.

It's an API-wrapper node from T8mars/ComfyUI_Seedance, so the usual deal applies: you need a key from api.seedance.nz, not a graphics card. And don't expect a free lunch - video super-resolution is the priciest per-call job in this pack.

How it works

The model rides the "legacy" compatible video endpoint (/v1/video/generations, not the newer /v1/videos). You give it exactly one source - either a local ComfyUI VIDEO (which gets uploaded automatically) or a public MP4 video_url - and the node stuffs that into metadata.video_url, submits the task, polls until it's done, then downloads and verifies the result. Downloads go through the pack's resilient chain: .part temp file, MP4 ftyp header check, and a fallback to system curl if the Python connection dies. That reliability layer is the part people actually hit, because result servers are slow.

The inputs that matter

• video_url - a public MP4 direct link. Leave empty if you're wiring in local video.
• input_video - the local ComfyUI VIDEO. This is the one you'll actually reach for 90% of the time: load a clip, drag it in.
• api_config - connect a Seedance API Config node, or rely on SEEDANCE_API_KEY.
• skip_error - on, it returns a placeholder video instead of killing the whole workflow. Nice for batch jobs.

Outputs are video (the upscaled VIDEO, ready for SaveVideo), video_url, task_id and response.

Where people get burned: the source must be 480P, 3–15 seconds - the model is trained for that input band, and the node won't negotiate. And it accepts exactly one source. Wire input_video and fill video_url and it'll refuse with a clear error. Keep them mutually exclusive.

Installing it

From ComfyUI Manager, search "ComfyUI Seedance", or:

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 ComfyUI, grab a key at api.seedance.nz/console, and connect the Config node. Only dependency in the whole pack is requests - no model files to babysit.

Common issues

• Source rejected - check duration (3–15s) and that it's actually 480P. Upscaling a 720P clip "just to be safe" is not the way; the model expects 480P.
• Both sources set - pick one: input_video or video_url.
• Slow downloads - the result CDN can be slow; the node's timeout budget (180s total) usually covers it, but big clips on a slow line can trip. If downloads fail on a proxy, check HTTP_PROXY/HTTPS_PROXY - a stale proxy env var is the classic culprit in this pack.

One honest caveat: for heavy jumps (256px → 1024px), the KB's own upscaling doc notes SeedVR2 beats FlashVSR on restoration quality. FlashVSR is the fast, free-your-GPU path; it's not the last word in detail-add.