🚀 Remove Background

Background removal is one of those operations where the old tools are fine until they aren't - and the moment you cut out a person with flyaway hair, you find out your u2net-based setup was never going to make it. This node runs BiRefNet, which is the model that displaced rembg's u2net as the default for hard edges: hair, fur, fabric mesh, semi-transparent material. You feed it an image, it returns a clean cutout and a mask, and that's the whole node. No separate model loader, no preprocessing chain, no manual thresholding.

One important context if you've been following along: BiRefNet is now bundled inside ComfyUI core (since May 2026), so you have a native path too. This node is the pack's all-in-one take - and honestly the main thing it adds over the core setup is a model picker and a background-color choice in one place.

How it works

The node vendors the ZhengPeng7 BiRefNet weights and downloads them on first use into ComfyUI/models/ComfyUI-YarvixPA/RemoveBackground/<model>/ via HuggingFace - which means the first run needs internet and takes a while. It then pads and resizes your image to the model's native resolution, runs inference in fp16 on your GPU, and post-processes the mask back to the original size.

Four models, and picking wrong is the most common mistake:

• BiRefNet (default) - the 1024×1024 general model. The right call for most images.
• BiRefNet_lite - Swin-Tiny backbone for tight VRAM. A step down in edge quality.
• BiRefNet_lite-2K - 2560×1440, for large sources on constrained hardware.
• BiRefNet_dynamic - handles arbitrary input resolutions natively. Use this for anything big - feeding a 4K image into the fixed 1024 model internally downscales it and throws away exactly the hair detail you wanted kept.

background_color picks what fills where the background was: transparency (RGBA output), white, or black. device defaults to auto. update_model (default off) forces a re-download if you think the weights are stale.

Outputs are image (the cutout) and mask (single-channel), which you can wire into a compositing or masking workflow directly.

Install

ComfyUI Manager (search "YarvixPA") or

cd ComfyUI/custom_nodes
git clone https://github.com/YarvixPA/ComfyUI-YarvixPA

restart. This is the one node in the pack with heavy real dependencies: the pack's requirements.txt pulls opencv-python and timm (BiRefNet's Swin backbone needs timm), plus torchvision which ComfyUI core already provides. Manager installs them.

Where people get burned

The big one: the model is loaded from disk on every execution. There's no caching - each run reconstructs the BiRefNet network, so don't drop this node inside a per-frame loop and expect real-time. It's a "run it once, save the result" node, not a per-batch engine.

Second: the transparency output is a flat RGBA image. If you save it with ComfyUI's SaveImage you get a PNG with alpha - good - but any downstream node that strips alpha (most do) silently flattens it. Grab the mask output too if you need to re-composite later.