Image Remove Background Model Loader

This is the loader half of the pack's background-removal pair: it builds a cutout network once, and you feed the result into the rembg_model input of Image Remove Background. The model dropdown is short and deliberate - it offers the BiRefNet variants and BEN2, not a grab-bag of fifteen networks you'll never use.

The reason you want a separate loader instead of the remove node just loading weights itself is memory discipline. Building a cutout network takes a moment and holds a few hundred megabytes, so the loader keeps the built network for the life of the process, and one loader can feed several Image Remove Background nodes - batch after batch - without rebuilding it each time. Run a folder of fifty images through background removal and that's a real saving.

Which model to pick

• BiRefNet General - the default, and it suits most pictures. BiRefNet is the current standard for background removal; it produces meaningfully sharper edges than the older u2net generation, especially on hair and semi-transparent materials. ComfyUI shipped it in core in 2026, which tells you how settled the recommendation has become.
• BiRefNet Portrait - trained on people, the choice when your subject is a person and you want portrait-optimized edges.

The one honest caution from the broader ecosystem: BiRefNet's quality drops if you feed it a huge image, because the 1024px weights downscale internally and throw away exactly the edges you were trying to keep. If you're cutting out a 2K+ photo and the result looks soft, the fix is usually resolution handling on the Image Remove Background side, not a different dropdown here.

Where the weights go

Weights go in ComfyUI/models/birefnet (and ComfyUI/models/ben2 for the BEN2 option), and they're downloaded there on first use when features.network: true is on in config.yaml. The pack's docs name the repositories for each variant. Offline or bandwidth-shy? Pre-place a .safetensors in models/birefnet named as the widget lists it and the loader finds it without the network flag ever being needed.

Building the network is cached for the process's life - that's the "loader" part of the name doing its job - so a second loader or a second remove node costs nothing extra.

Installing

WASRembgModelLoader ships in WAS Node Suite v3. ComfyUI Manager → search WAS Node Suite v3, or:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui

Restart after; needs ComfyUI 0.14.0+ and Python 3.10+. The pack itself installs nothing, but this node is a model consumer, so plan on either turning on features.network for the one-time download or dropping weights in models/birefnet yourself. That's a deliberate architecture: v3 downloads nothing unless you ask, unlike the old v2 suite that pulled a whole dependency stack at install.

The minimal working graph is: Image Remove Background Model Loader → Image Remove Background → a preview or save. One loader, one remove node, and a cutout with edges that don't look like they were cut with scissors.