PD aitookit training redux

PD aitookit training redux (PD_AitoolkitTrainingRedux) sorts a training dataset folder into two piles: images that have a matching caption file, and images that don't. That's it, and for dataset work it's surprisingly handy. It's named after ai-toolkit because that's the training pipeline the author was prepping data for, but it'll tidy a folder for any LoRA or fine-tune trainer that expects image+caption pairs.

Why this matters

Dataset curation beats every training knob you can turn - that's one of the few things that survived the SDXL-to-Flux trainer changes intact. A folder that's half-labeled, half-not will silently train garbage captions on the unlabeled half, or just error out when the trainer expects a .txt for every image. Running this node before you start is the "check your work before you spend hours training" step.

How it works

It scans source_folder, and for each image it checks whether a same-named text file exists (matching text_extension, default .txt). Pairs get copied to paired_folder_name (default paired_files); images or text files without a partner go to unpaired_folder_name (default unpaired_files). Source files stay put - it copies, it doesn't move - so nothing is destroyed if you misconfigure it. It outputs a message string summarizing what got sorted, which doubles as a sanity report. image_extensions (default .png,.jpg,.jpeg,.webp) sets what counts as an image.

The inputs that matter

• source_folder - where your dataset lives. Required, typed by hand.
• paired_folder_name / unpaired_folder_name - the two output subfolders it creates inside the source.
• image_extensions / text_extension - the file-type rules. Defaults are right for the standard image.png + image.txt convention.

Installing

Part of the 7BEII/Comfyui_PDuse pack. ComfyUI Manager → search "Comfyui_PDuse" → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/7BEII/Comfyui_PDuse
cd Comfyui_PDuse
pip install -r requirements.txt

Restart after. No models; the pack's dependencies (Pillow, numpy, opencv-python, etc.) are all this node needs.

Gotchas

Two things to know before you run it. First, matching is by exact filename - 1_R.png pairs with 1_R.txt, and 1_T.png without a 1_T.txt lands in unpaired. If your captions use a different naming scheme, this node won't magically reconcile it. Second, it copies rather than moves, which is safe but means a big dataset gets doubled while you sort - run it on a copy or accept the disk cost. And the message output is a plain string, so wire it into a text preview or Show Text node if you want to read the report; it's an output node, so nothing downstream depends on it.