Remove background (full)

The simple Remove background node is the pack's "just do it" button. This is the "fine, let me have every knob" version. Remove background (full) runs the same model inference but exposes everything: pre-processing size, interpolation, foreground colour estimation, fill colour, mask threshold, and it hands back the mask, depth, and edge outputs the simple node hides. The README is upfront about the one cost - it consumes more RAM - but if you're doing real compositing work, this is the node you'll live in.

The inputs that matter

• model (SET_REMBG) + images (IMAGE) - the loader output and your input, as everywhere.
• width / height (default 1024) - the size images are scaled to before the model runs (doesn't change the output size). Wire the loader's train_w/train_h here for the model's sweet spot.
• upscale_method (default bicubic) - scaling in and out. Bicubic is usually right.
• blur_size (default 90) / blur_size_two (default 6) - diameters for the coarse and fine Gaussian blurs used by the Approximate Fast Foreground Colour Estimation. This is what makes the edges of a cutout take on the subject's true colour instead of bleeding the background's colour into hair. The advanced example workflow shows swapping this for FMLFE (fine-level estimation), which the author says is better but far heavier.
• fill_color (boolean) + color - with no background image connected, this fills the background with a solid colour instead of leaving it transparent. The color field accepts hex (#RRGGBB), comma-separated RGB, rgb()/hsv()/hsl(), or ~140 named colours. Black by default.
• mask_threshold (default 0) - force the mask to hard 0/1 above a confidence level. Leave at 0 to keep soft alpha (right for matte models).
• batch_size - video/batch throughput; 3060-class GPUs are better at 1.
• depths (optional) - external depth maps for PDFNet (otherwise auto-computed).
• background (optional) - an image to composite behind the subject. It's scaled to match images, and for video it can be another video with the same frame count. Leave empty for RGBA transparency.
• out_dtype - float16 for long video jobs to halve RAM.

Outputs

• images (IMAGE) - the cutout: RGBA with transparency, background-replaced, or colour-filled.
• masks (MASK) - the estimated foreground masks.
• depths (MASK) - depth map (PDFNet only).
• edges (MASK) - edge map (DiffDIS only).

Install and the usual gotchas

ComfyUI Manager search "Remove Background (SET)", or:

cd ComfyUI/custom_nodes
git clone https://github.com/set-soft/ComfyUI-RemoveBackground_SET
pip install -r ComfyUI-RemoveBackground_SET/requirements.txt

ComfyUI 0.3.48+, seconohe >= 1.0.6 (fresh pip install git+https://github.com/set-soft/seconohe.git if it errors).

Two things that trip people up. First, RGBA output isn't supported by every downstream node in ComfyUI - the pack's own examples keep the image and mask separate, which is the robust pattern for anything compositing-related. Second, the author's honest framing: there's no universal best model, so this node's value is being able to compare - the 09_Compare_Models example runs ten models through this pipeline side by side. When a cutout looks bad, the fix is usually not more knobs but a different checkpoint; the knobs just let you wring out the last bit of edge quality.