Alpha Composite

Fun fact about this pack: the README only documents two nodes, but the source ships four, and Alpha Composite is the one nobody told you about. It's a general-purpose layer compositor - think "photoshop layer stack inside your graph" - that takes a foreground, a background, and blends them with independent offsets, scaling, blur, and opacity. It's not strictly part of the crop-and-paste story, but it's the glue you reach for when you want a soft-edged element placed over an image without opening a raster editor.

Why you'd want it in a workflow: detail stitching is all about controlled placement, and sometimes the controlled part isn't a crop box but a composited layer - a watermark, a corrected face region, a foreground object feathered into a scene. This node does the Porter-Duff "over" composite (foreground over background, alpha-aware) that ComfyUI's core nodes oddly make you hand-assemble.

How it works

It splits each input into RGB plus an alpha channel - either from the image's own alpha (if it's already RGBA) or from an optional background_mask / foreground_mask you feed in. Then, in order: the foreground is scaled (fg_scale, 0–10, 1 = no scaling), both layers get a separable Gaussian blur applied to their color and alpha (so edges feather into transparency rather than cutting off), opacity is multiplied in, and both are placed onto a canvas sized to the background using bg_offset_x/y and fg_offset_x/y (each ±8192). Finally the "over" operation combines them.

The inputs that matter most for a beginner:

• background and foreground - the two images.
• fg_scale - resize the foreground without touching the background.
• fg_blur / bg_blur - 0–200, feather the layer edges. Small values (2–8) hide seams; big values give a soft glow look.
• fg_opacity / bg_opacity - 0 to 1, exactly what they sound like.
• fg_offset_x / fg_offset_y - nudge the foreground around the canvas.

Outputs are image and mask. One honest warning: the image output is 4-channel RGBA, not the 3-channel IMAGE most ComfyUI nodes expect. Either split the alpha back off before wiring downstream, or use the mask output for anything that needs a mask.

Installing it

Same pack as everything else here - ComfyUI Manager (search Detail Stitch) or:

cd ComfyUI/custom_nodes
git clone https://github.com/mancusog/comfyui-detail-stitch.git

Restart and it appears under the "Detail Stitch" category. No requirements file, no models, just torch.

Where people get burned

• The 4-channel output surprise. You'll get a "size mismatch" or a black-ish image somewhere downstream if you don't account for the alpha. Split it or feed nodes that accept RGBA.
• Blur before placement. The blur runs before offsetting, so a blurred edge doesn't smear across the offset - which is right, but if you expected a Photoshop-style drop shadow, this isn't it.
• No documentation. Since the README never mentions this node, the only ground truth is the code. The defaults are sane (everything starts neutral), so start there and adjust.