BiRefNet Background Removal (CRT AutoDL)

Before Pixal3D can turn your photo into a 3D asset, it needs a subject. The whole pipeline is tuned to reconstruct the thing in the middle of the frame - backgrounds confuse the depth estimation, mess up the camera FOV guess, and leak noise into the geometry. So the first real step of a Pixal3D workflow is cutting the subject out, and the model the community actually reaches for for that job is BiRefNet. This node is a one-click version of that step.

CRTAutoDLPixal3DBiRefNet downloads birefnet.safetensors into ComfyUI/models/background_removal/ on first run and hands it to ComfyUI's native background-removal loader. Same AutoDL pattern as the rest of the pack: fixed file, auto-download with a console progress bar, then delegate to the core node so the actual loading logic is ComfyUI's own. No inputs at all.

BiRefNet deserves its reputation. The Bilateral Reference Network was originally built for dichotomous image segmentation, but it became the default background remover because it keeps hair, fur, and semi-transparent edges far better than the older u2net lineage - and ComfyUI shipped it in core back in May 2026. For Pixal3D input prep, "good enough edges" isn't enough; a stray halo of background becomes a shelf of phantom geometry, so BiRefNet's edge quality is exactly what you want before the back-projection starts.

The one output

bg_model - a BACKGROUND_REMOVAL model object. Wire it into ComfyUI's native RemoveBackground node (which applies the model and hands you the cutout image), then feed that into ImageCropToMask and on to Pixal3DConditioning. The conditioning node's tooltip even spells out the intended input: an image cropped to the mask with pad_factor=1.1.

Install

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes
pip install -r requirements.txt

Restart ComfyUI, or use ComfyUI Manager and search CRT-Nodes. You need a current ComfyUI - this node calls comfy_extras.nodes_bg_removal, which is core. If RemoveBackground isn't in your node search, update ComfyUI.

Gotchas

• First run downloads the model. It's roughly a gigabyte; the CRT AutoDL progress bar shows in the console. One time, and it lands in models/background_removal/ where ComfyUI's own Load Background Removal Model node can reuse it.
• You get one fixed BiRefNet checkpoint, not the whole family. The pack downloads the standard 1024px general weights. If you're cutting out very large images (north of ~1500px), the BiRefNet_HR 2048 variant is a better fit and you'd load that through the native node instead - this CRT node is for the standard path.
• Don't background-remove a LoRA training set with this. The KB's most repeated finding: stripping backgrounds from training images teaches the model to emit blank backgrounds. That advice has nothing to do with this node specifically, but it's worth remembering before you repurpose it.

The setup is a little counterintuitive - a node whose only job is to download a file and hand you a model socket - but it removes the single most error-prone step of assembling a Pixal3D graph by hand. You get the right BiRefNet file, in the right folder, loaded through ComfyUI's own machinery, every time.