Load Image (Preserve Alpha)

Here's a subtle ComfyUI trap: load a PNG with transparency through the built-in Load Image, and you get the pixels with the alpha ripped out into a separate MASK output - the transparent regions are just gone from the image itself. For most workflows that's fine. For compositing, it's a problem, because you often want the RGBA data as one image with alpha as a real fourth channel, the way PNG actually stores it.

UC_LoadImageWithAlpha is Core Load Image plus exactly one extra thing: an IMAGE_RGBA output that carries the alpha as a fourth channel alongside the normal IMAGE and MASK outputs. Drop-in compatible with Core's loader - same image picker from the input folder - but you get the transparent version when you need it.

The three outputs

• IMAGE - the flattened RGB image, identical to what Core's loader gives you. Wire this into anything that doesn't care about transparency.
• MASK - the alpha as a mask, again matching Core behavior.
• IMAGE_RGBA - the image with alpha preserved as the fourth channel. This is the new thing, and it's what a compositing node that wants real transparency eats.

One input, image - a combo dropdown listing files in ComfyUI/input, with upload support, exactly like the Core node. If you can use Core's Load Image, you already know how to use this one; it's the same muscle memory with an extra output dangling off the right side.

When it matters

The moment you're building layered/compositing pipelines. This pack's own compositing family (UC_LayeredBackgroundComposite, UC_BackgroundRemovalPreserveAlpha, and friends) deals in cutouts and RGBA - if you're feeding it a pre-made transparent PNG, the alpha is the whole point, and losing it to a flat Load Image silently turns your subject into a black rectangle. It also matters for any workflow that composites PNG stickers, sprites, or render passes with transparency. There's a reason the author positions it as "Core Load Image, but you also get the RGBA."

Install

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

Restart (or ComfyUI Manager → search "ComfyUI-UtilsCollection"). Pack deps are opencv-python and typing-extensions - nothing extra, because it's a loader, not a model.

The one thing to watch: not all downstream nodes accept a 4-channel image, and some will silently drop or mangle the extra channel rather than error. If your composite suddenly looks wrong, check that the node receiving IMAGE_RGBA actually knows what to do with it - otherwise use IMAGE for the flat path and save RGBA for the compositor. Simple node, specific job, and it does it exactly right.