Transparent Background Ultra (RBS)

Want a subject on a transparent background, saved as an actual PNG with alpha, no compositing gymnastics? This is the pack's one-stop version of that. It wraps the transparent_background library - the same RMBG-based engine behind lyra's ComfyUI_TransparentBackground - and turns an image into an RGBA cutout plus its mask in one pass. It's fast, it's a known commodity, and it's about the simplest thing in the RemoveBackgroundSuite: image in, transparency out.

The honest caveat before you get excited: this is the older, less refined end of the cutout world. RMBG is decent and quick, but if you're cutting out hair on a 4K product shot, the pack's own BiRefNetUltra node with the HR weights will give you the better edge. Reach for this one when you want a fast, good-enough cutout, or when you want the mask input trick without the big model download.

The dropdown that looks broken (and isn't)

The model dropdown is populated by scanning ComfyUI/models/transparent-background/ for .pth files at node-creation time. An empty dropdown means you haven't dropped a checkpoint in there yet - it's not a bug. Grab the checkpoints from the transparent-background project (they're named after their modes):

• ckpt_base.pth → the "base" model
• ckpt_base_nightly.pth → "base-nightly"
• ckpt_fast.pth → "fast"

Drop whichever you want into that folder, refresh the node, and the dropdown fills in.

The gotcha that will actually bite you

Here's where people get burned. The pack's requirements.txt does not list transparent-background, but the node does from transparent_background import Remover at runtime. So the first execution dies with ModuleNotFoundError unless you install it yourself:

cd ComfyUI/custom_nodes/ComfyUI-RemoveBackgroundSuite
pip install transparent-background

Install the pack, install that, drop in a checkpoint, and you're actually running. If you came from the BiRefNet node expecting everything to be bundled, this one is the exception - it's the one node in the suite with a missing piece.

The knobs and the trick

• max_megapixels (default 2.0) - same downscale-for-inference-then-resize-back logic as the rest of the pack. Raise it for large sources.
• device - cuda or cpu.
• auto_use_original_mask (default on) with the optional mask input - the same trick as the BiRefNet node: if a valid mask arrives, the model is skipped entirely and your mask becomes the alpha. Great for turning a hand-fixed mask into a transparent image without re-running inference.

Outputs and wiring

Two outputs: image (RGBA - wire it to Save Image and make sure PNG is selected) and mask, which you can route into the pack's Mask Process Details node to soften the edges if the raw cutout looks harsh.

Install and wrap-up

Same pack, same install as the others: clone into custom_nodes, pip install -r requirements.txt, restart. Manager finds it by "RemoveBackgroundSuite". Add transparent-background, add a checkpoint, and you're done - a genuine two-node transparency pipeline that a beginner can wire in five minutes.