StabilityAI Suite - Remove Background

Remove Background does exactly what it says and nothing else. Drop in an image, get back the same image with the background cut out and an alpha channel. One input, one output, and the entire node is a thin wrapper around Stability's stable-image/edit/remove-background API endpoint.

How it works

There's no cleverness to explain: your image is POSTed to the endpoint, and the response is an RGBA PNG with the subject isolated. No prompt, no mask, no settings. The segmentation model does its thing server-side, and the node hands you the result.

Inputs and output

Just image. The output is a single IMAGE - save it as PNG (or WebP) to keep the transparency. JPEG will flatten the alpha onto whatever background color the encoder assumes, which is a classic way to lose the whole point of the operation.

Install

It's part of the same pack, so the usual two paths:

# ComfyUI Manager → search "ComfyUI StabilityAI Suite" → Install
# or manually:
cd ComfyUI/custom_nodes
git clone https://github.com/florestefano1975/ComfyUI-StabilityAI-Suite

Restart ComfyUI, set your API key (SAI_API_KEY env var, sai_platform_key.txt, or the per-node override), and you're done. There are no models to download and no pip requirements - the pack only uses libraries ComfyUI already bundles.

When you'd reach for it - and when you shouldn't

This is the most honest "convenience node" in the pack. If you're cutting out a handful of images and don't want to install a local background-removal stack, it's one node and you're done. But if you're batch-processing product shots, every image is a paid API call, and a local model becomes obviously cheaper - the standard options (BiRefNet and the RMBG-family nodes) run free on your own GPU once downloaded.

The other thing worth knowing: this is a fork of Stability's official nodes by florestefano1975, who also maintains Portrait Master and Prompt Composer, so the code follows the same clean shape as the official releases. The pack itself has little Reddit presence - you won't find a crowd troubleshooting it, but there's genuinely nothing to troubleshoot: one input, one call, one image back.