遮罩批次合成 (Ref)

SCAIL-2 has a specific appetite: it wants a reference image and a colored support mask that says which pixels are the character and which are background. Getting those two into agreement - same subject, same background logic, same ordering - is where most first-time SCAIL-2 workflows quietly go wrong. MaskCompositeRef is the pack's dedicated preprocessor for that, sitting between your reference images + masks and the Wan SCAIL To Video (Multi Ref) node.

It's not a general compositor. It's built for SCAIL-2's two generation strategies, switched by replacement_mode:

• Replacement mode (replacement_mode = true) - for moving a character out of their old background. Your masks have a black background (RGB < 10/255 = background, colored = person). The node auto-drops any fully-black mask frames (and the matching source image), blackens the background of the remaining images, and ignores the backgrounds input entirely. Output: person cut loose from their surroundings, ready for replacement.
• Animation / action-transfer mode (replacement_mode = false, default) - for moving a character's motion onto a new background. Masks have a white background. The node composites the person onto the first image of your backgrounds batch, and any extra background images get inserted into the sequence (between the second-to-last and last frames). process_first_frame / process_last_frame control whether the first/last source images participate.

In both modes remove_background is the kill switch for the compositing itself: off, and it only does frame-dropping/size alignment, passing the rest through.

What you set

• images - the reference sequence, order must match masks one-to-one (the node does not validate ordering - that's on you).
• masks - the colored mask batch. Not connected → passes through, with an all-black (replacement) or all-white (animation) mask output.
• backgrounds - only meaningful in animation mode.
• The four booleans above.

Outputs are images and masks, both sized to match, ready for Wan SCAIL To Video's reference_image and reference_image_mask inputs. The mask background rules line up exactly with the pack's Create SCAIL-2 Colored Mask node, so the two compose cleanly.

Install

Ships in user2318/ComfyUI-CustomNodeKit. ComfyUI Manager: search "CustomNodeKit". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/user2318/ComfyUI-CustomNodeKit.git
cd ComfyUI-CustomNodeKit
pip install -r requirements.txt

then restart. No extra models - it's numpy/torch math and a LANCZOS resize.

The gotcha

The failure modes are all "garbage in, masks-out-of-sync." Wrong background color: feed a white-background mask into replacement mode and the node thinks the whole frame is background - this is the #1 reason a SCAIL-2 reference comes out as a floating smudge. Out-of-order pairs: since the node doesn't validate ordering, a shuffled images batch gives you a character wearing someone else's mask. And in replacement mode, if every frame is all-black the node returns an empty batch - which sounds like an error but is actually the pack's intended "skip this segment" signal, so a downstream conditional should handle it rather than panic. Match the mask background to the mode, keep the ordering identical, and this node does its one job beautifully.