🧹BRIA RMBG

This is the node that does the actual work in the two-node BRIA pack: feed it the model from 🧹BRIA_RMBG Model Loader plus an image, and you get back the subject cut out against transparency, alongside a clean mask. It's the closest thing in this pack to a one-click background remover, and on a solid subject against a busy background it's genuinely good.

The inputs and outputs that matter

Two required inputs, exactly what it says on the tin:

• rmbgmodel - the model object from BRIA_RMBG_ModelLoader_Zho (the loader has to run first)
• image - the IMAGE tensor you want cut out

And two outputs:

• image - the cutout, an RGBA image with the background made transparent
• mask - a single-channel MASK you can wire into anything that takes a mask

That's the whole surface. No strength slider, no prompt, no "keep the top 30% of background" option.

How it works

The pack's own code tells the story. For every image in your batch, it converts the tensor to a PIL image, downscales to 1024×1024, normalises with mean 0.5 and std 1.0 into the range the network expects, and runs the U2Net-style BriaRMBG model. The first of the model's side outputs is a soft alpha map. That map gets upscaled with bilinear interpolation back to your original resolution, stretched to full range, and used as the alpha channel: the original pixels are pasted onto a fully transparent canvas, masked by the cutout. So the alpha is real soft-edged data rather than a jagged threshold, though it's still a hard segmentation, not true matting - a veil or a wine glass will confuse it.

Two details matter in practice. First, resolution: inference always happens at 1024×1024 regardless of your input. A 4K product shot is internally shrunk, cut, then blown back up, so fine edge detail like flyaway hair gets softer than the model could manage. That's the fixed input size doing its thing, not a bug you can tune away.

Second, the pack handles batches. The README's headline feature since V1.5 is feeding multiple images - or a stack of video frames - through at once and getting every one back cut out. That's exactly what the author's own demo does: SVD-generated video frames, RMBG on each, background gone.

Where it fits, and how it's aged

Released February 2024, this is an early single-model wrapper, and it shows its age against the modern field. ComfyUI has shipped BiRefNet natively since May 2026, and the fat GPL pack ComfyUI-RMBG bundles half a dozen remover models behind one set of nodes. The honest take: RMBG-1.4 sits in the same league as the old rembg/u2net baseline - excellent on people and products with a contrasting background, noticeably weaker than BiRefNet on hair, fur, and genuinely semi-transparent material. The community comparison from the era said much the same: not obviously better than SAM, and it gives you no control over what gets isolated - foreground or nothing.

So reach for this pack when you want one dead-simple cutout and no model zoo. If you're doing production cutouts on tricky edges, BiRefNet (native now) or InSPyReNet is the upgrade path.

Installing and wiring

Same drill as the loader article, briefly: git clone the pack into custom_nodes (or ComfyUI Manager → search "BRIA RMBG"), restart, then manually download model.pth from briaai/RMBG-1.4 into the pack's RMBG-1.4 folder. No pip install - the deps are all ComfyUI staples. The one gotcha: if the model file is missing, the loader errors out before this node ever runs.

Common issues

• The RGBA output surprises people. The image output has four channels. Save Image writes a proper transparent PNG, and most compositing nodes handle it, but RGB-only nodes can flatten or choke on the alpha. If you only need the selection for masking or inpainting, take the mask output and skip the image entirely.
• Soft edges on big inputs. The 1024 fixed inference resolution again. Downscale deliberately and know what you gave up, or switch to a higher-res model.
• It cut out the wrong thing. RMBG isolates the salient foreground, full stop. No box, no text prompt, no "the one in the middle." For picking a specific object, that's SAM/GroundingDINO territory.
• License. Non-commercial for the open weights, under BRIA's custom license. Read it before shipping anything.

The loop is short and predictable: load the model, drop in an image, read image or mask. For a beginner who wants a transparent PNG without installing five node packs, that's a feature, not a limitation.