Image Alpha Composite

You cut a subject out with a background remover, it came out as an RGBA image with real transparency, and now you want to drop it onto a new background - without threading a separate mask node through the graph. That's this node: it layers image_2 over image_1 using image_2's own alpha channel as the blend weight. Where image_2 is opaque, it fully replaces image_1. Where it's transparent, image_1 shows through. Where it's partially transparent - soft edges, hair, glass - you get a proper blend instead of a hard cutout line.

How it works

Standard alpha-over compositing: for every pixel, the output is a weighted mix of image_2 and image_1, weighted by image_2's alpha at that pixel. Fully opaque alpha (1.0) means you see only image_2; fully transparent (0.0) means you see only image_1; anything in between blends proportionally. This is the same math behind "paste a PNG with transparency onto a background" in any image editor - just as a graph node instead of a manual layer stack.

The inputs and outputs that matter

Two required inputs, both IMAGE:

• image_1 - the base, or background.
• image_2 - the layer going on top. Its alpha channel is what drives the composite, so it needs real transparency data to do anything useful here.

One output: IMAGE, the composited result.

How to install it

Part of the Art Venture pack. ComfyUI Manager: search comfyui-art-venture, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/sipherxyz/comfyui-art-venture

Restart ComfyUI. No model downloads - it's pixel math.

Common issues & troubleshooting

image_2 just fully overwrites image_1, no blending at all. image_2 doesn't actually have an alpha channel - it's opaque everywhere, so full-opacity compositing is exactly correct, even if it's not what you wanted. Some loaders and nodes flatten or drop alpha along the way; confirm whatever produced image_2 (a background-removal node, a URL loader) is actually preserving transparency rather than silently discarding it.

Cutout edges look hard and jagged instead of blending. That means image_2's alpha is binary (fully on/off) rather than graduated. A segmentation model with a hard threshold produces exactly that; a matting-capable model or a blurred mask edge produces softer alpha and a cleaner composite here.

Mismatched resolutions. Both images should match in size before compositing - resize one to the other first if they came from different generation or loading steps.

Wanted a Photoshop-style blend mode instead (multiply, color, etc.), not a straight overlay. This node does one thing: alpha-over compositing. For the "recombine luminance from one image with color from another" trick, that's ColorBlend, a different operation entirely.