zhenzhen-bria-video-background-removal-v3-fal

Background removal is the most commoditized operation in this whole ecosystem - the KB calls it "a problem that looks solved until you hit hair, fur or anything semi-transparent." For stills, you've got a dozen local options (BiRefNet, InSPyReNet, rembg) and ComfyUI even added native support. For video it's a different story: cutting a subject out of every frame, keeping the matte stable across motion, and doing it with an alpha you can composite on - that's where local tooling gets fiddly and slow. This node just delegates the whole thing to BRIA's bria/video/background-removal/v3 endpoint through the Zhenzhen API. Upload a clip, get back a cutout.

Why reach for it instead of a frame-by-frame BiRefNet loop? Temporal consistency. A per-frame model flickers on hair and semi-transparent edges because every frame is decided independently; a video-specialized matting model keeps the edge stable across frames, which is exactly what makes the difference between a cutout you can use and a cutout that shimmers. If you're compositing a talking head onto a fake background, or building green-screen-less product video, this is the fast path.

How it works

The flow is the standard Zhenzhen FAL pattern: give it a video (via the video input or a video_url), it uploads it to the workshop's file service, POSTs to the proxy, polls until the job finishes, then hands back a VIDEO plus the hosted video_url. You don't even need to connect a video - video_url is the required field and is ignored whenever a video input is wired in.

Inputs and outputs that matter

• background_color - the interesting one. Default Black, with Transparent first in the list. If you're compositing later, pick Transparent (or one of the 9 solid colors if the video host you're feeding can't handle alpha).
• output_container_and_codec - ten options. Default is webm_vp9, which is the right call for transparent output in most pipelines (mp4 can't carry alpha). If you're feeding a compositor that wants ProRes, mov_proresks is there. Need a quick shareable loop? gif.
• preserve_audio - on by default, keeps the original track. Turn it off if the clip's audio is garbage and you'll replace it anyway.
• video_way - upload (default) sends your file; video_url lets you point straight at a hosted clip and skip the upload.
• poll_interval / max_poll_attempts - the timeout budget. Default is 600 polls × 6s = an hour, which is a lot of patience for a short clip; you can trim it.
• skip_error - on, failures return a blank video and log instead of killing your batch.

Outputs: video (the cutout clip as a VIDEO tensor, ready for save or compositing), video_url (hosted file), and response (raw JSON).

Installing it

One pack, one clone:

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

Restart ComfyUI, or install "Comfyui-zhenzhen" via ComfyUI Manager. No model downloads - the matting happens on BRIA's servers.

Common issues

Pack-level gotchas: you need a Zhenzhen account with credit; the overseas API can throw 443 from some networks (VPN with TUN mode); 500s on first run are usually upstream - re-run.

For this node specifically, the two traps are container and color. Transparent video only works if the container supports alpha - mp4_h264/mp4_h265 cannot carry it, which is why the default is webm. And if your cutout comes back with a hard black edge, you almost certainly chose Black as the background color instead of Transparent; solid colors are for when you want to key it out later, not for actual removal. One more: video_url is required, so if you only have a local file, wire the video input or the node will sit there waiting for a URL that never comes.