RembgByBenAdvanced

RembgByBenAdvanced is RembgByBen with the training wheels off. Same two-step pipeline - BEN mask, then blur-fusion foreground color estimation - but nothing is hidden behind defaults, which means you can fix the cases where the one-click version isn't good enough.

Under the hood, and it's worth knowing because it explains the interface: this class literally inherits from GetMaskByBen and BlurFusionForegroundEstimation, and its single function calls both in order. Every input you see belongs to one of those two steps.

• Segmentation side: width / height (1024) - the preprocess resolution the model sees; does not affect the final output size. upscale_method (bilinear) - interpolation used for the preprocessed image and the mask resize. mask_threshold (0) - set above 0 to zero out low-confidence pixels and scrub floating noise off the mask.
• Foreground side: blur_size (91) and blur_size_two (7) - the coarse and fine blur radii for the Photoroom fast-foreground-estimation defringing. fill_color (false) and color (0–16777215, i.e. 0xFFFFFF) - flip fill_color on to replace transparency with a solid RGB background instead.

Outputs: image (the RGBA cutout, or solid-background if you filled) and mask.

When do you actually need the knobs? A few honest cases:

• Small or low-res subject: drop width/height to 512. The 1024 default is overkill and just slows you down.
• Hair with a grey fringe that won't die: lower blur_size from 91 toward ~31 so the halo removal doesn't smear the fine strands.
• A mask full of specks: nudge mask_threshold up to ~0.02–0.05 and watch the preview. Too high and you'll eat the soft edges.
• Product mockup on a coloured backdrop: fill_color on, color set to your brand hex.

Install and model situation are shared with the whole pack: clone, pip install -r requirements.txt, restart ComfyUI, and put BEN_Base.pth or BEN2_Base.pth in ComfyUI/models/rembg/ben/ - nothing auto-downloads, and the filename decides BEN vs BEN2 (see LoadRembgByBenModel).

My take: run RembgByBen first. If the output is clean, you never needed this node. The moment you hit a fringe you can't explain, switch to this and start turning dials. The defaults are sane, but they're not magic - this is where a background-removal workflow actually gets tuned.