LC Image Compare 🔎

You've got two batches of images and you need to flip between them - version A of a workflow versus version B, before and after an upscale, two seeds of the same prompt. The standard answer is an image comparer node, and if you've used the popular ones you know the pain point: feed them a large batch and the top of the node becomes a row of clickable numbers that wraps, overflows, and shoves the image area around so hard the node physically jumps every time the batch size changes. LC Image Compare (class LCBatchImageComparer) is built specifically to not do that.

The source is upfront about it: it's a drop-in replacement for the rgthree-style Image Comparer that keeps the image area position and size constant. Instead of the growing number list, it uses a compact dropdown to pick which pair you're looking at, and it always reserves a fixed-height placeholder so the image doesn't jump - even on a single-image run. If you've ever resized a batch and watched a comparer reflow itself into a mess, you know exactly why that matters.

Mechanically it's a viewer node, not a processor. It takes image_a and image_b (both optional, both IMAGE), lets you flip between them, and outputs images - the pair you were comparing - so the node can sit in the middle of a graph and pass the images along, or just serve as a review endpoint. It's marked as an output node, meaning it runs at the end of a queue, and it inherits the pack's preview behavior. The pairing is positional: index 0 of batch A compares with index 0 of batch B, and the dropdown walks the pairs.

That "output the images" bit is worth underlining because a comparer that only previews is a dead end. Here you can wire the images output into a Save Image or a further processing chain, which turns the comparer into both a review tool and a pass-through - handy when you want to inspect a batch before committing it to disk.

Where people get caught: the two batches need to be the same length, or at least you need to know which is shorter - comparison is index-based, so a 12-image A against a 4-image B gives you four sensible pairs and eight comparisons into empty space. And because it's a viewer, it only shows what you feed it; if you're comparing the effects of a node, the standard trick is to fork the graph and run your variant on one branch so both inputs hit this node with the same resolution.

This is one of the pack's "Performance" settings exceptions too - the README notes that LC Image Compare deliberately isn't affected by the hide-FX-previews performance options, because a compare node you can't see is useless.

Install is the standard LC123 clone:

cd ComfyUI/custom_nodes
git clone https://github.com/lonecatone23/ComfyUI_LC123_nodes

Restart, no pip deps. If you A/B test anything with a batch size bigger than five, the dropdown layout alone is worth the switch.