Seedvr Upscale Video (fal)

SeedVR2 is ByteDance's restoration model, and it's the current default upscaler in the local community - it adds detail without reinventing the subject, which is exactly what you want. There's a catch, though, and it's the reason this node exists: SeedVR2 on images runs on a 6GB card, but SeedVR2 on video is a datacenter-GPU workload. One person put it plainly - "I tried video upscaling with it but I quickly realized I'd need datacenter class GPUs." So this is one of the cleanest cases for an API node in the whole pack: you get the best video upscaler without renting a GPU yourself.

What it actually does

Nothing runs locally. Your video goes to fal.ai, fal runs SeedVR2's video upscale, and you get a higher-resolution clip back - no VRAM, billed per job. SeedVR2's one-step design is fast for what it is, and its strength is faithfulness: it sharpens and rebuilds detail while leaving the content alone, rather than hallucinating a new picture the way heavier restorers can.

The inputs that matter

• video or input_video_url: the source. Wire a VIDEO into video, or paste a link into input_video_url - you provide one of the two.
• upscale_factor (required, default 2): how much to enlarge when in factor mode. 2× is the safe default.
• upscale_mode: factor (multiply the resolution by upscale_factor) or target (aim for a fixed resolution instead).
• target_resolution: 720p, 1080p, 1440p, or 2160p - used when upscale_mode is target.
• output_quality: low → maximum. Higher looks better and costs more compute.

noise_scale (default 0.1) controls how much fine detail the model synthesizes - nudge it up carefully, since too much starts to invent texture. output_format picks the container (.mp4, .webm, .mov, .gif) and output_write_mode trades file size against speed. The output is a STRING video_url; download it with the pack's Fal Save Media from URL node.

Installing it

Easiest: ComfyUI Manager → search ComfyUI-fal-API → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
cd ComfyUI-fal-API
pip install -r requirements.txt

Restart ComfyUI. No model download; the dependencies are fal-client and a fal API key (in config.ini under [API], or the FAL_KEY env var). If you're feeding a VIDEO input, a recent ComfyUI is needed for the native video socket.

Where people get burned

Two things carry over from SeedVR2's known behavior. First, it lifts contrast - the local community's most specific complaint is that outputs can read as more AI-generated than the input because of that contrast bump. If your clip looks over-punched, that's why; keep noise_scale modest. Second, SeedVR2 preserves what's there but doesn't repair heavy degradation - it's an enhancer, not a magic restorer, so a genuinely mangled source stays mangled.

On cost, video upscaling is compute-heavy by nature, and maximum quality at 2160p is the expensive corner - test on a short trim before committing a long clip. Errors come through as fal's real text, and the Windows Portable import fix is .\python_embeded\python.exe -m pip install fal-client.