I kept running into the same annoying step before training a LoRA: a folder full of images, matching .txt caption files, and me bouncing between Finder, Preview, and a text editor trying to keep everything clean.
It works, but it gets old fast. Especially when you’re checking for missing captions, fixing trigger words, doing the same small tag edits over and over, or trying to make sure you didn’t skip an image.
So I made a small native macOS app for it.
It’s called LabelSmith. It just opens a local dataset folder and lets you review images, edit the matching sidecar caption, find missing captions, search through the set, and do some basic batch caption cleanup.
Basically the app I need before starting the actual training step.
Right now it expects a simple flat dataset folder like:
image_001.jpg
image_001.txt
image_002.png
image_002.txtIt’s still early, but it’s open source here:
https://github.com/atlasnotch/LabelSmith-MacOS
If anyone here prepares their datasets locally on macOS or anywhere else, and has a dataset prep workflow they hate, I’d be curious what’s missing. I’m mainly building this around the small annoying tasks that make caption cleanup slower than it needs to be.