This project aims to help with the creation and management of LoRa training datasets. Scroll down to the bottom of the page for a feature overview.

This is still in beta - please report any bugs you find Pull requests are welcome. Currently everything is just cobbled together

Roadmap:

• Refractor tag management
• Add single-image tag overrides [mostly done]
• Add superresolution/upscaling for small images

Known issues:

• No files/folders are ever deleted, leading to clutter/orphaned images
• Deleting an enite folder can break the step
• When an image is open and you save the dataset, it will throw an error and half of the files are left where they are.

(optional) create a venv first:

python -m venv venv
venv\Scripts\activate

install the requirements:

pip install -r requirements.txt

start either by running start.bat or manually using:

python webserver.py

(see python webserver.py --help for launch arguments)

Access the webui on the following URL: http://127.0.0.1:8080/

If tagging/cropping/etc is too slow, try run pip install onnxruntime-gpu, but keep in mind that this will use some of your VRAM.

Running this script is recommended to get all features of the webui.

using start.bat already downloads all dependencies by default

It will gives you the option to download the following files:

• danbooru-tags.json and gelbooru-tags.json from github gist or catbox.moe.
  • You also have the option to scrape the tags from the site directly.
• cropper.js and cropper.css from Cloudflare/cdnjs.

Clear your browser cache between updates. It tends to leave the old scripts/css loaded

The folders created are meant to be used as follows:

• 0 - raw - raw images from the internet / screenshots
• 1 - cropped - cropped images (1:1 aspect ratio)
• 2 - sorted - images grouped by quality / topic / etc
• 3 - tagged - .txt or .json files containing autotagger output
• 4 - fixed - pruned tags in .txt format.
• 5 - out - scaled down images and pruned tags - point your training script here
• datasets - all your datasets are saved here

Some of these images/videos might be outdated. There's also in UI tooltips. If something breaks just open an issue here on Github.

• Save / load datasets you're working on
• Avoid having to change training folder, just point your training script at the 5 - out folder and load the right dataset
• Write notes for yourself

• Crop images in your browser
• Edit already cropped images
• Duplicate image - crop two separate parts
• Quickly set the cropped area, copy it from the previous image
• Keyboard shortcuts

• This is a good "baseline" to work from
• You can go back and quickly fix what it misses
• It can detect multiple subjects.

• add categories
• quickly sort multiple images, captcha style
• hit detection can be janky

• Don't want to do it manually? Set the tags and sort automatically.

• A bit slow but does not take any vram, doesn't influence training.
• The output isn't realtime, I think it can do about 1 image/sec on my 11gen i5

• Prune useless tags from the autotagger
• Normalize tags
• Quickly blackist/whitelist tags
• Replace tags on all images
• Edit rules, test the effects

• Scale images to required training resolution