Load Image Alpha

The stock ComfyUI loader quietly destroys the alpha channel. LoadImage opens your PNG, converts it to RGB, and the transparency is just gone - you get a square image where your cutout used to be. That's the gap LoadImageA fills: it's the same familiar upload-picker, but it can keep the alpha as a real RGBA tensor, and it hands you a mask and the file path as bonuses.

It comes from comfyui-missed-tool, a small "a few tools that I wanted" pack by ifmylove2011. The name undersells it a little; this node is genuinely the thing you reach for when you build workflows around transparent assets.

How it works

Under the hood it's plain PIL plus a touch of torch. It opens the file, and the single switch on the node decides the rest:

• rgba_mode on → the image is converted to RGBA and the tensor keeps all four channels.
• rgba_mode off (the default) → it converts to RGB, alpha dropped. At that point it's just LoadImage with an extra string output.

It then reads the alpha band, if the file has one, and emits mask = 1 - alpha. Note the inversion: this matches ComfyUI's own mask convention, where white means transparent / to be removed. If the file has no alpha band at all, you get a silent 64×64 zero mask instead.

The inputs and outputs that matter

One input matters: rgba_mode (BOOLEAN, default false). The other, image, is the same upload/choose dropdown you already know.

The three outputs are where the value is:

• image - the IMAGE tensor, RGBA if you flipped the mode.
• mask - derived from the alpha, in ComfyUI's white-means-transparent convention.
• image_path - the resolved file path on disk. This one's the sleeper hit: wire it into a text or file node to load a matching caption, a settings JSON, or whatever sits next to the image.

A classic setup: LoadImageA feeds an RGBA cutout straight into a compositing or overlay step, while image_path pulls the sidecar file that tells you what to do with it.

Installing it

Easiest route is ComfyUI Manager - search for comfyui-missed-tool and install. Manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/ifmylove2011/comfyui-missing-tool
cd comfyui-missing-tool
pip install -r requirements.txt

Then restart ComfyUI. The pack's dependencies are torch, numpy, Pillow - all of which ComfyUI already ships, and there are zero model downloads. Install is boring, which is exactly what you want from a utility pack. The nodes land under a "missed-tool" section in the menu. One naming quirk: the README spells the clone URL "comfyui-missed-tool" while the repo is "comfyui-missing-tool"; GitHub redirects the old spelling, so either works.

Where people get burned

• The default kills your alpha. If you just plug in the node and don't touch rgba_mode, you've silently recreated plain LoadImage. If your output has square corners, flip the toggle - that's the number one gotcha.
• The mask is inverted on purpose. 1.0 means transparent here, matching ComfyUI. Don't "fix" it with an Invert Mask unless you know your downstream node wants the other polarity.
• No alpha band → junk mask. A 64×64 zero tensor comes back silently. Chain that into a mask-based operation and you'll get a confused result with no error to tell you why.

Newer ComfyUI builds did add their own "Load Image (RGBA)" loader, so on a fresh install you have a choice. LoadImageA's edge is still the one-stop shop: RGBA image, mask, and path out of the same node - handy in a batch or a workflow you're tidying up.