A collection of useful scripts related to AI workflows
no warranty provided, use at your own risk.
install with pip install -r requirements.txt
to make batch masked images (recommended settings):
python batch_create_masks.py --path C:/path/to/training/images --out C:/path/to/mask/outputs --edges --min-total-area=1 --contain --draw-contain
some options: use python batch_create_masks.py --help to view all options
--include-textfile: includes a .txt with the text read from the image
--corners: only create masks touching a corner
--edges: only create masks touching an edge or corner
--min-area: the minimum area for a single bounding box. Below this the bounding box is not drawn, as a 1-100 percentage.
--min-total-area: The minimum area that must be occupied by all bounding boxes, as a 1-100 percentage.
--max-area: the maximum area between all bounding boxes. If the total area of bounding boxes is greater than this amount, the mask is skipped, as a 1-100 percentage.
--contain: trace a bounding box around all combinations of bounding boxes; then remove bounding boxes that fall outside the largest bounded area within --max-area. This will help eliminate detections that are not located near other detections.
--draw-contain: optionally use with --contain, draws a bounding box around the traced area.
--contain-under-min: optionally used with --draw-contain, only draws contain if the detected boxes have less than the total area already.
--use-color=False (defaults to true) converts image to grayscale before performing OCR
--use-binary converts the image to black/white before performing OCR. Try --use-color=False before trying this option
for img2img batch masking:
DPM++2m SDE Karras, 10 steps
Resize by 1
mask mode: INPAINT NOT MASKED!!! (areas to clear are covered with black boxes). If WHITE areas are meant to be removed, choose INPAINT MASKED.
masked content: fill
For keep tokens, pick the token to advance to the front of the captions. Regex matchs first instance of foo to end of string, and boosts:
python boost_caption_tokens.py --path ./texts --regex "foo.*$"
otherwise, just boost a single token:
python boost_caption_tokens.py --path ./texts --regex "foo,"
python random_gather.py --path=/path/to/image/folders --outpath=/path/to/output/folder --ignore-ends