实例商品裁切为背景图

If you run a marketplace shop, you know the drill: every listing needs a square product image on white, product centered, nothing else in frame. KSProductsToBackground is that step, automated. Feed it segmented product instances plus their masks from the SAM pipeline, and it crops each one tight, centers it on a white or transparent square, and hands you a clean batch - plus a manifest telling you which instance became which output.

What it does

It's the "make it listing-ready" node, and it's the payoff of the whole detect→segment chain:

• segmented_images + masks - the batch from KSSAMSegmentByBoxes (or KSGroundedSAMSegment). Each mask defines one product instance.
• background_mode - white or transparent output.
• output_size (default 1024, range 256–4096) - the square canvas edge length. Bigger = more detail, bigger files, slower.
• margin_percent (10) - padding around the product. 0 fills the canvas edge-to-edge; higher shrinks the product in the frame.
• sort_order - top_to_bottom numbers outputs reading down-then-across; left_to_right reads across-then-down. Matters if the order needs to match a catalog sheet.
• edge_feather (0.8) and the same mask-cleanup knobs as KSApplyMaskBackground (mask_threshold, min_mask_component_area, mask_close_radius, mask_expand_pixels, fill_mask_holes).

The interesting optional input is original_image. When you connect it, the node crops each product from the original photo using the mask as a guide instead of from the (already-blackened) segmented image. The tooltip says it plainly: the original usually keeps more real texture and color. If your segmented crops look washed out, this is the fix.

There's also prefer_grouped_product_masks (default on): when several small parts sit under one big mask, it keeps the big product mask whole instead of fragmenting into pieces. Off, and you get more separate components - which you might want for multi-part products.

The outputs

• products - the batch of centered square crops.
• manifest - JSON recording product count, output index, source instance index, and each original crop box. Genuinely handy if you need to trace outputs back to inputs.
• product_count - how many made it through.
• original_background - one white/transparent full-canvas image with all detected products kept together. That's the "group shot" version, complementing the individual squares.

Installing

With the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/kongshan4219/ComfyUI-Kongshan-Nodes

restart ComfyUI. Local, no keys; needs the SAM models only if you're using the upstream segmentation nodes in the same pack.

Gotchas

• The "cutout" is only as clean as the mask. This node crops by mask, so a feathered or sloppy mask produces a halo. Use mask_threshold up and edge_feather low for crisp catalog edges.
• 1024 default × many products adds up. If you're producing 20 listing images, that's 20 × 1024² PNGs. Drop output_size for drafts, raise it only for finals.
• The 1688 shape of it. prefer_grouped_product_masks and the square-white-centering are tuned for Alibaba-style marketplace listings, which is the author's world (this pack is extracted from their private e-commerce workspace). If that's your world, this node is genuinely the fastest path from a messy photo to a clean catalog sheet.

Young, tiny, unknown pack - but this is one of its two or three genuinely differentiated nodes. There's nothing else in the typical ComfyUI install that does "SAM masks → centered square white-background product batch + manifest" in one node.