Image Comparer (PixNodes)

You've got two samplers running the same seed against two different prompts, or one prompt against two checkpoints, and now you need to actually look at both before you pick a winner. The Image Comparer is the node for that: two image inputs, one draggable split view, and - this is the part people miss - it never writes a thing to your output folder.

It's a straight port of rgthree's Image Comparer, which itself has been the community's go-to A/B tool since forever. The pack's source says so: ported from rgthree, with all the external dependencies stripped out. That matters because it means the node works on a bare ComfyUI install with nothing else feeding it.

How it works

Under the hood it subclasses ComfyUI's built-in PreviewImage, so it has no output ports at all - it's a display node, full stop. Give it up to two images and it saves them to ComfyUI's temp directory (that's the type: "temp" flag, not output) and hands them to the frontend, which draws the pair and lets you drag a divider back and forth across the canvas. The labels it shows are F for the foreground image and B for the background, which maps to the two input names.

The drag is the whole point. Pixel-peep the hair, the hands, the sky - the divider is instant, no re-running, no waiting for a file to open in another app.

The inputs

Only two, and both optional:

• foreground_image (IMAGE) - the "after" side, conventionally.
• background_image (IMAGE) - the "before" side.

Wire your two VAE Decode outputs into them, or a before/after pair from an upscale workflow. Leave one unconnected and it just shows the one that's there. That's the entire interface.

Install

It ships in the Comfyui-PixNodes pack. Easiest route is ComfyUI Manager: search "PixNodes", hit Install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/pixixai/Comfyui-PixNodes

then restart ComfyUI. Heads-up: the README's manual-install snippet contains a copy-paste error and tells you to clone ComfyUI-AlignLayout - wrong repo, use the URL above. There are no model downloads and no heavy dependencies for this node; the pack's requirements (openai, Requests) only exist for its LLM API nodes, so this one costs you nothing extra.

Where people get burned

Because it's a display node with zero outputs, you can't chain anything off it. If you need the compared image downstream - say, to save the winner - keep a normal Preview Image or Save Image wired in parallel. Also don't expect keyboard shortcuts or fancy tiling; this is the stripped-down port, not the full rgthree widget. It's the node you drop in while iterating, not the one that finishes the job.