Unified Background Replace

Most background-replacement nodes handle one subject at a time. This one takes up to fifty foreground images, cuts them all out, and independently places each one over a single shared background - all in one queue, with per-foreground outputs you can inspect and composite yourself. If you've ever manually re-ran a cutout node fifty times to build a group shot, you know why this exists.

What it is

A batch background-replace node. It uses an internal BiRefNet model to segment each foreground, cleans up the masks, resizes each cutout relative to the shared canvas, positions it with two shift sliders, and composites. The important design choice: each foreground gets its own output image and mask, rather than everything stacked into one final composite. That means you can check every placement before committing, and the compositing control stays in your hands.

How it works

The pipeline per foreground: BiRefNet produces a soft mask → mask_threshold (0.5) decides what counts as solid foreground → border_cleanup_width strips weak predictions at the source edges → artifact_cleanup_radius (opening) removes small specks → gap_fill_radius (closing) fills cracks and holes → feather_radius softens the edge inward while keeping the interior opaque.

Sizing and placement are the clever part. foreground_scale is defined as the fraction of the background's shortest side occupied by the foreground's longest bound - 0.9 default means the subject nearly fills the canvas; values above 1 overscale and crop at the edges. long_axis_shift and short_axis_shift slide it along the background's long and short axes, each from -1 (one end) through 0 (center) to 1 (the other end). workspace_padding (0.5 default) lets placements sit partly off-canvas by up to 25% per axis, which is how you get natural edge-cropped compositions instead of everything floaty in the middle.

The inputs that matter

• background - the single shared canvas.
• foreground_images - the autogrow batch of subjects to cut out and place.
• foreground_scale - size relative to the background's short side.
• long_axis_shift / short_axis_shift - placement.
• mask_threshold - how confident the model must be to keep a pixel.
• background_removal_model (optional) - plug in a Core background-removal model to override the internal BiRefNet.

Outputs: images (the composed results) and masks (the placement masks for each output) - both IMAGE/MASK, sized to match the number of foregrounds.

Installing it

Part of ComfyUI-UtilsCollection:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart ComfyUI (or use Manager, search "ComfyUI-UtilsCollection"). No model download step on your end - BiRefNet is pulled from the Comfy-Org repo the first time the node runs. The pack's own deps are minimal and it needs a recent ComfyUI, since BiRefNet support and the background-removal model type are Core features from 2026.

Common issues

BiRefNet is the right default for hair and fur - that's exactly what it beats the old u2net at - but fine hair at low mask_threshold can still fuzz out, so raise it if you get halos, lower it if you lose strands. Watch foreground_scale: above 1.0 you're intentionally cropping, which surprises people the first time a subject's head vanishes. And with 50 foregrounds, memory adds up - the node is batch-happy, but your GPU isn't infinite. If the internal model's first run is slow, that's the one-time download, not a hang.