Instant Reference Load Images In Folder

A boring little loader that turns a folder of reference images into the IMAGE batch the rest of this pack eats. You give it a directory path, it reads the images in it, and out comes a stacked tensor plus a list of the file paths. That's the whole job, and it's genuinely useful because feeding 12 reference images into Instant Reference LoRA by hand-dragging them every time is exactly the kind of friction this pack exists to remove.

Point it at a folder, wire images into a training node, and the workflow becomes "drop files into a directory, hit run." Same job the built-in Load Image does one image at a time, but in bulk.

Inputs

• folder_path - a string, the directory to scan. This is the only one you must set.
• recursive - whether to also scan subfolders. Default off.
• max_images - a cap on how many images to load. Default 0 means no limit. If you're training, this is your accidental-dataset guard: 5 or 10 keeps you honest.
• resize_to_first - default on. If the folder has mixed sizes, it resizes everything to match the first image (LANCZOS). Turn it off and mixed sizes become an error instead.

Outputs

• images - the IMAGE batch, all frames stacked into one tensor. Wire this into Instant Reference LoRA or Instant Reference LoRA Train.
• image_paths - a string with the resolved paths, one per line. Handy for a quick "which files did I actually load" check, or as text you can stash in a workflow note.

Install

ComfyUI Manager → search "Instant Reference" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/cstria0106/comfyui-instant-reference

Notes

This node does no captioning, no training, and no model downloads - it's pure image IO, so it shares none of the pack's first-run heaviness. Two small things worth knowing: the folder path is a plain filesystem path (no ComfyUI-specific folder picker widget here), and if resize_to_first is off, mismatched dimensions will raise an error that tells you exactly that - at which point you either turn it on or fix your folder. For reference training, resizing to the first image is usually the right call anyway; a consistent dataset is a happy dataset.