Image Compare (HTML)

Every serious workflow eventually hits the same wall: you generated twenty variations of a LoRA, a checkpoint, a cfg sweep, and you need to actually look at them side by side before picking a winner. In core ComfyUI that means Save Image nodes, a file explorer, and a lot of alt-tabbing. Image Compare (HTML) is the node that collapses that whole ritual into one run: it takes your batch and writes an interactive HTML gallery you open in the browser - grid, before/after slider, opacity overlay, synced zoom and pan, even a loupe. It's the best node in Kinburg-Nodes, and honestly the reason to install the pack at all.

It's also the node in the pack that leans hardest on the pack's wider ecosystem. It's designed to sit at the end of an accumulator-driven sweep - wire it the outputs of Get Accumulator (images), (captions), (prompts), (texts) - so a batch of runs becomes a browsable page with each image's caption, full generation prompt, and per-image generation time underneath. The settings_data input (from the pack's Generation Info Filter) renders each run's actual sampler parameters under the image, which is the thing you genuinely need when you're comparing a 12-step recipe against a 30-step one and you've forgotten which was which.

The headline trick is the reference input. Wire in a baseline - the source of an upscale, or the fp16 output when you're checking a quantized model - and the node computes SSIM + PSNR for every image against it and adds a "Similarity" sort to the page. That's a real objective tiebreaker on top of your eyes, and it turns the node from a gallery into an evaluation tool. (No reference? reference_index uses one image in the batch as the baseline instead; -1 turns metrics off.)

Only images is strictly required. The rest of the inputs are sockets, not widgets - wire captions (one per line), prompts (one block per image, separated by a --- line), times, judge_data (a Vision LLM Judge's results, if you want AI verdicts rendered per image). For output: images_captioned is a list of images with captions drawn on (so mixed-size inputs stay separate rather than being forced into one batch), html_path is the file, and url is a clickable served link - open the page straight from the node.

How it's saved

By default it writes a portable folder - <prefix>_<datetime>/ with a light index.html and an images/ subfolder using relative links. Zip it, share it, open it offline, it all works. Flip embed_images on and you get one self-contained .html with every image inlined as base64 - single file, much larger. Both open from the url output. Your review state - hide, star rating, tags, comments - persists in browser localStorage, so you can come back to a sweep the next day and still see which one you rejected.

Installing

It's part of Kinburg-Nodes (ComfyUI Manager → search "Kinburg-Nodes", or git clone https://github.com/Kinburg/Kinburg-Nodes into custom_nodes and restart). Per the README, Image Compare needs nothing extra beyond ComfyUI's bundled deps. The full sweep setup - accumulators, Generation Info - comes with the same pack, so it all slots together.

Where people get tripped up

The auto_collect toggle defaults to on and quietly re-wires every accumulator in your graph right before the workflow queues, so a Set node you just added or muted is picked up without a click. If you ever see images appearing on the page that you didn't think you'd collected, that's the culprit - flip it off and use the manual 🔌 Collect All button. And don't forget the page's visibility toggles: captions, prompts, settings, metrics and judge sections each have their own on-page switch, so a "cluttered" comparison is usually just one toggle away from clean.