Concurrent Submit | zhenzhen-pixelcut-video-background-removal-fal

Here's the thing about background removal: locally, it's commoditized to death - BiRefNet, rembg, a hundred nodes, all free. But do it on video and the local story changes completely. Per-frame matting of a multi-second clip on your own GPU is slow and fiddly. This node sends the whole video to Pixelcut's background-removal service through fal's queue and hands you back a clean cutout video with the background already stripped. That's a genuinely useful cloud job, not just a local model wearing an API costume.

It's also a video node, which means it runs in the shared video pool (10 workers, not 30), and it returns a COMFLY_VIDEO_FUTURE task. Collect with Concurrent Collect Videos (ComflyConcurrent_Video_Await), which takes up to 10 tasks and returns video_1…video_10 in slot order.

The inputs

• video - the input clip (VIDEO type from your video nodes), or video_url if you have a hosted URL. The video_url tooltip says it's ignored when video is connected, so wire the tensor when you can.
• video_way - upload (send the clip to their server) vs video_url (pass a URL through).
• background - what to put behind the subject: transparent, black, white, green, blue, magenta, or custom. Transparent is the default and the usual goal - green/blue are for keying workflows downstream.
• custom_r / custom_g / custom_b - the RGB triple for custom backgrounds.
• output_format - the good stuff: auto, webm_vp9, mp4_h264, mp4_h265, mov_proresks, mov_h265, mkv_h264, mkv_h265, mkv_vp9, or gif. ProRes for post, h.264 for web, gif when you must.
• poll_interval / max_poll_attempts - default 600 × 6s = 1 hour timeout.
• api_key and skip_error - the usual.

Install

Pack-standard - ComfyUI Manager → search Comfyui-zhenzhen, or:

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

Restart, add your token (ai.t8star.org overseas; api.seedance.nz for the domestic discount) to api_key.

Things worth knowing

• Upload size is your limit. You're shipping a whole video file to a server. Long 4K clips will be slow to upload and slow to process; keep the input short if you're iterating.
• Transparency in every container isn't free. If you pick mp4_h264, you're getting the background you selected - you only get real alpha in formats that support it (webm, mov, mkv). That's how video containers work, not a node bug.
• This is the video pool. Don't expect the 30-way image concurrency here; 10 concurrent video jobs is what the pack gives you, and each one is hammering the same pool as every other video node in the pack.