Captions To Prompt List

If you've ever needed all the captions from a training dataset as one searchable blob - say, to find which tags dominate, or to sample lines for a prompt experiment - you know how tedious it is to concatenate fifty .txt files by hand. CaptionsToPromptList walks a directory tree, reads every caption file, and hands you a single combined string plus a suggested output filename. It's a dataset utility, not a creative node, and it's exactly the kind of thing DJZ-Nodes throws in alongside its fancier stuff.

How it works

The mechanism is simple and honest: it uses os.walk to recurse through directory_path, and for every .txt file it reads the lines and appends them to one list, then joins everything with newlines. Outputs are combined_captions (STRING) and output_filename (STRING) - the latter is just the folder's name plus .txt, e.g. my_dataset.txt, meant to save you from typing the same name into SaveText.

The reload toggle is the one clever bit. ComfyUI caches node results; with reload off, the node only recomputes when the input directory_path changes. Flip reload on and it computes an MD5 hash of every caption file's contents, so it re-runs when any caption changes - useful while you're actively editing captions and the graph should pick up the edits.

What it's for

Two main workflows:

• Dataset QA - dump combined_captions into a text viewer or a wordcloud node to eyeball tag distribution before training a LoRA. The KB's lora-training essay is firm that caption quality is the hidden variable in LoRA results, and this is the cheap way to inspect what you actually have.
• Prompt source material - wire the output into a text-splitter or into the Zenkai prompt nodes in the same pack to sample real captions as prompt fragments.

It ignores non-.txt files with a console message (except .png, which it skips silently), and it prints the path of every file it processes - noisy in the log, but reassuring for large runs.

Installing

It's part of DJZ-Nodes (MushroomFleet / Drift Johnson). Grab the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/MushroomFleet/DJZ-Nodes
cd DJZ-Nodes
pip install -r requirements.txt

Or "DJZ-Nodes" through ComfyUI Manager. The node itself is pure stdlib (os + hashlib), so no extra dependencies.

Gotchas

• Point it at the folder containing the caption files, not a single file. It recurses, so a nested dataset layout works fine.
• It reads every line of every caption as a separate entry - a multi-line caption becomes several lines in the output. Usually that's what you want; just don't be surprised by the line count.
• With reload on and a huge dataset, the hashing pass adds startup time on every run. Leave it off unless you're actively editing captions.