🖼️+📝 Load Text-Image Pairs (List)

This is the batch version of this pack's Load Text-Image Pair (Single) node. Instead of picking one image/caption pair out of a dataset folder, it hands you the entire dataset at once as parallel lists - every image, every matching caption, rotated so the pair your seed points at comes first.

How it works

Same basename-matching logic as the Single node - image1.png pairs with image1.txt (or whatever extension text_format_extension is set to) - but the output is a full set of lists rather than one item. seed still matters here, but as a rotation point rather than a strict pick: it's where the returned lists start from, cycling through the rest of the set after that. force_reload bypasses whatever's cached from disk, which you want after editing caption files and re-running without restarting ComfyUI. limit_count caps how many pairs come back at all (0 means everything).

The inputs and outputs that matter

• seed - the rotation starting point into the dataset.
• folder_path - the dataset folder, used only when image_input/text_input aren't connected.
• force_reload - forces a fresh read from disk instead of using a cached result.
• limit_count (optional, default 0) - cap the number of pairs returned; 0 = no cap.
• image_input / text_input (optional) - connections that take priority over folder_path.
• text_format_extension (optional, default txt) - the caption file extension to match.

Outputs: image_list, string_list, image_path_list, image_filename_list, and total_count.

How to install it

Via ComfyUI Manager: search ComfyUI-mnemic-nodes, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/MNeMoNiCuZ/ComfyUI-mnemic-nodes

Restart ComfyUI. No model downloads or extra dependencies.

Common issues & troubleshooting

Mismatched counts between images and captions. Same caveat as the Single node, just at scale - if a caption file is missing for one image in the set, that pair still comes back but with an empty string in string_list at that position rather than the whole load failing. Worth a quick scan of the lists (or total_count against a manual folder count) before trusting a large dataset is fully paired.

Editing captions and not seeing the changes. This is what force_reload is for - leave it off during normal use and it's fine, but if you've just hand-edited a .txt file and re-queue without seeing the update, that's a caching issue force_reload fixes.

Loading way more than you meant to. For a large dataset, set limit_count while you're still testing a downstream pipeline - there's no reason to pull ten thousand images into memory just to check that your batch captioning or metadata-stripping logic works on the first fifty.