Blind Comparer

You generated eight images and need to pick the best one, but you know which seed produced which frame - so you're not really choosing on merit, you're choosing on expectation. Blind Comparer is the fix: a tournament bracket that shows you two images at a time without telling you which is which. You vote Left or Right, the bracket advances, and a champion emerges - chosen on what it actually looks like, not what you think it should look like.

It's the kind of node that sounds like a gimmick and turns out to be quietly valuable, because preference bias is real and it's expensive. When you're A/B-ing samplers, schedulers, LoRA strengths, or seeds, the whole point is that the image should win, not your memory of the config. This node enforces that.

How it works

• image_1 - the first image input, which is always present.
• Additional slots auto-expand as you connect more images - wire any number in.
• Images pair up in a knockout bracket; odd counts get a bye (the node description calls this out explicitly, so don't worry if your count is odd).
• Each round you vote Left or Right on the two shown, and winners advance until one champion is left, displayed at the end.

It's an output node with no outputs - it's a judging tool, not a passthrough.

When you'd use it

• Seed selection. Batch-generate a few seeds of the same prompt and pick the winner blind.
• Parameter A/B tests. "Does the cache-accelerated run actually look worse?" is a question you should answer with your eyes, not your bias about the technique. Blind Comparer is the honest way to settle it.
• Hair-splitting quality checks. When two upscalers or two denoise values look "basically the same," a blind bracket forces a real call, and it's surprisingly decisive.

Install

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

Restart ComfyUI, or install via ComfyUI Manager (search "WtlNodes"). No models, no heavy deps.

Gotchas

• "Blind" means you still see the images - just not the labels. If your thumbnails have the seed burned into a filename or a watermark overlay from another node, that defeats it. Feed it clean outputs.
• It's interactive. This node pauses your workflow for a human to vote; it doesn't run unattended. Great for a focused judging session, wrong for a fully automated batch.
• More images = more rounds. A 16-image bracket is 15 votes. That's the cost of rigor - and honestly, if you can't be bothered to make 15 decisions, you probably didn't need 16 candidates.