RAM Image Compare

RAMImageCompare is the "which one is actually better?" node. Feed it two images and it shows them side by side in the node preview with a divider you can drag across - one image on the left of the line, the other on the right - so you can compare detail at the exact same spot in both. And it does it all in RAM, writing nothing to disk. This is the node that turns "let me save both and open them in an image viewer" into "flip the slider, done."

Inputs are dead simple: image_a and image_b, plus a compare_mode dropdown with two options. slide gives you the draggable divider - drag it left and right to reveal each image under the other. click is a toggle: click and hold to flip between the two views. slide is the default and, for most people, the more useful one - it's how you actually A/B fine detail like film grain intensity, sharpening amounts, or the difference a LoRA strength made.

Mechanically, the node grabs the first frame of each input tensor, encodes both to PNG in memory, and ships them to the frontend over ComfyUI's websocket in a single message. The frontend then does the slide/click rendering. No save, no output/ folder pollution, no temp files to clean up later. If you're iterating on a comparison - tune, rerun, compare - the lack of disk I/O is the whole point.

The gotcha is the one this pack keeps tripping people on: this is an output node with no wire outputs, and it previews in the browser. Run it headless or via API and there's no browser to show anything in - the node still "runs," but you get nothing to look at. It's a frontend tool; keep it in workflows you run from the UI.

It's part of the WtlNodes pack (Scorpiosis0/ComfyUI-WtlNodes). Install via ComfyUI Manager by searching "WtlNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Scorpiosis0/ComfyUI-WtlNodes.git

then restart ComfyUI. Deps are numpy, scipy, and pillow; no models, no keys.

Worth noting: it only ever compares the first frame of each input. If you feed it batches or videos it's still comparing frame zero, so this is for stills - which is what the description implies and what most A/B workflows want. For the "pick a winner from many" problem, the pack also has a Blind Comparer node; RAMImageCompare is the two-image, side-by-side version.