Image Blend by Mask

Compositing has a boring-but-essential step between "put image A over image B" and "it looks real": blending the two so the boundary isn't a hard cut. UC_ImageBlendByMask does that with a mask plus your choice of blend mode - and unlike a one-mode paste, the mask is resized and broadcast for you, which removes an entire class of "these dims don't match" errors.

The shape is classic: a destination (the base you keep) and a source (what you blend in), with a mask deciding where the source shows through. Where the mask is zero, you get the destination; where it's one, you get the blend; in between, a mix. blend_percentage (default 1.0) scales the whole thing down when you want a partial effect.

Blend modes

mode defaults to add, and it's a real Photoshop-style list: add, color, color_burn, color_dodge, darken, difference, exclusion, hard_light, hue, lighten, multiply, overlay, screen, soft_light. A quick map for the ones you'll actually use: screen lightens and is great for glows or light leaks; multiply darkens and is the standard for texturing/shadows; overlay and soft_light punch up contrast; difference is a debugging/inversion tool; color transfers the source's hue without its luminance. Everything clamps to 0–1, so runaway add won't blow out to white garbage.

The inputs that matter

• destination / source - the two images. If resize_source is off (default), dimensions and batch size must match, or you get a clear error.
• resize_source - turn it on to have the source bicubic-resized to the destination's size instead.
• mask - optional; resized to destination dimensions and singleton batches broadcast across the image batch, which is the "connect one mask to a batch of frames" convenience.

Single output: blended_image.

Installing it

ComfyUI Manager → search "ComfyUI-UtilsCollection", or:

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

Restart. Pack deps are opencv-python and typing-extensions, no model downloads. Newer-API pack - keep ComfyUI reasonably updated.

Where it bites: forgetting resize_source and connecting mismatched sizes gets you an error (fine, at least it's explicit), but connecting a mostly matching pair with the mode set to add can silently over-brighten - check blend_percentage before blaming the mask. And for hard-edged masks, the pack's ModifyMask/blur route is your friend; this node blends colors, it won't soften a jagged mask for you.