Magnific Remove Background

One input, one job: hand it an image, get back a cutout with the background gone, plus a mask you can reuse. Outputs are image, mask, creation_identifier and metadata - and the optional folder from Magnific Save To, without which the result lands in your Personal project.

It works, it's simple, and it's the one node in this Magnific family where the argument for using it is weakest - because the same pack ships a local background remover that needs no account, no credits and no upload.

What it does and doesn't expose

Genuinely nothing to configure on the node itself. No threshold, no blur, no model choice - the model picker lives on the service side, and auto behaviour is Magnific's business. That's the design: it's a thin wrapper over their tool. Compare it to the pack's Background Remover (MEC) node, which gives you a model selector (RMBG-2.0, BiRefNet general, BiRefNet portrait), a threshold that swings between soft and hard alpha, mask_blur, and an invert for keeping the background instead of the subject - plus a foreground output delivered premultiplied.

If you only need an opaque cutout, the local node is the better first stop. It's free, it's fast, it's offline, and BiRefNet's edges are competitive with anything hosted. Two caveats worth knowing, both from the pack's own documentation: RMBG-2.0's weights are licensed by BRIA for non-commercial use only - the local node's other backends (SAM, BiRefNet, ViTMatte) are Apache-2.0 or MIT - and hair-level matting is a different job from background removal, which is what the matting nodes are for.

So when do you use this one?

When you're already in Magnific and want the result saved into a project folder alongside everything else that run produced, or when you want their specific cutout on a hard image where the local models disappoint. Comparing the two on your own worst image is the only honest test, and it's cheap to run.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git

Or "CustomNodePacks" in ComfyUI Manager, then restart. The Magnific nodes come with this pack - they're a port of Magnific's vendor pack magnific-comfyui 0.7.0, so don't also install the vendor zip or you'll have duplicate node names. Sign in once via the Magnific menu or the Magnific Save To node; credentials go to ~/.magnific/comfyui_auth.json and never into your workflow.

If you want the local remover, it's already there - no models to download manually, no extra Python packages beyond what the pack asks for (opencv-python, scipy, safetensors). Do check before installing those; ComfyUI ships its own numpy, torch and Pillow, and the README is explicit that a blanket pip install -r requirements.txt can break your install.

The trade

Every call uploads the image to Magnific's servers, costs credit, and comes back through their pipeline. For a shot you're doing a hundred variations of, that's a hundred uploads where a local model would have been free. For a handful of hero images where you want the best cutout and you're paying for Magnific anyway, press the button and stop overthinking it.

The mask output is the sleeper here: wire it into this pack's local masking nodes - grow, feather, threshold, composite - and you get the hosted model's edge decision with local control over the matte.