TS Remove Background

If you've cut out a subject and lost the flyaway hair, you've met the difference between u2net-era background removal and BiRefNet. TS Remove Background wraps BiRefNet - the model the background-removal doc calls the default recommendation as of mid-2026 - and adds the practical conveniences: a model picker, resolution control, mask blur/offset, temporal smoothing for video, and three outputs (the cutout, the mask, and a mask preview image).

This is the "isolate the subject" node for product shots, compositing, feeding clean alphas into other tools, or just killing a busy background before an img2img pass. It's a download-on-first-use model node, so unlike the pack's pure-math nodes it needs a moment to fetch weights - but it's a one-time cost.

How it works

BiRefNet (Bilateral Reference Network) was actually built for dichotomous image segmentation, not background removal - it just happens to be exceptional at it. The architecture's trick is its two-module design: a localization module that places the object from global semantics, and a reconstruction module that rebuilds the mask against two references at once - hierarchical image patches and gradient maps. That gradient supervision is why hair and fine edges survive; it pushes attention onto fine detail.

The node fetches weights from the ZhengPeng7 HuggingFace repos on first use, into models/BiRefNet/<variant>/. The model dropdown matters more than most: HR-matting is the default and the one to keep for high-res input; general is the classic BiRefNet; portrait is tuned for people; DIS is the dichotomous-image-segmentation weight. process_resolution (default 1024, snapped to multiples of 64) sets the working resolution - this is the VRAM dial.

The inputs that matter

• image and enable - feed the image; toggle enable off to pass it through untouched.
• model - the weight variant. Default HR-matting is right for most people.
• process_resolution (with use_custom_resolution) - the working resolution. Drop it if you're OOM.
• mask_blur and mask_offset - edge softening and boundary expansion/shrink. A couple pixels of blur kills the "sticker" look on hard edges; mask_offset positive expands (rescuing wispy edges), negative trims.
• background / background_color - Alpha (transparent) or a solid color via a COLOR widget.
• precision - auto picks bf16 on Ampere+ GPUs, fp16 elsewhere, fp32 only for diagnostics. Leave it on auto.
• temporal_smooth - for video: median3 (default) kills single-frame flicker, median5 is stronger but lags two frames at clip boundaries, ema_causal is causal with no lag, off for stills. This is what stops a video cutout from "boiling."

Outputs: IMAGE (the cutout - transparent or on the chosen background), MASK (the alpha), and MASK_IMAGE (the mask as a viewable image). Wire IMAGE to a saver/compositor and MASK into anything that needs alpha.

Installing it

comfyui-timesaver via ComfyUI Manager ("Timesaver") or:

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-timesaver
cd comfyui-timesaver
python -m pip install -r requirements.txt

Restart ComfyUI; weights download on first run.

Gotchas

The classic failure is an out-of-memory at high process_resolution - the troubleshooting section of the README points straight at this node's resolution control. Drop it to 768 or 512 and rerun. Second: if a model variant fails to download, the pack falls back through mirror sources (the ZhengPeng7 repos have occasionally gone gated or archived), so a transient failure usually resolves itself on retry. And the honest caveat from the background-removal doc: no model wins everything. BiRefNet is the quality default for hair and semi-transparency, but for anime or line art you may find other weights better - the model dropdown exists precisely so you can A/B without leaving the graph.