Automatically create masks for inpainting with Stable Diffusion using natural language.
txt2mask is an addon for AUTOMATIC1111's Stable Diffusion Web UI that allows you to enter a text string in img2img mode which automatically creates an image mask. It is powered by clipseg. No more messing around with that tempermental brush tool.
This script is still under active development.
Simply clone or download this repo and place the files in the base directory of Automatic's web UI.
From the img2img screen, select txt2mask as your active script:
Mask Prompt field, enter the text to search for within your image. (In the case of the topmost screenshot, this value would be 'business suit' and the prompt box at the top of your UI would say 'sci-fi battle suit.')
Mask Precision field to increase or decrease the confidence of that which is masked. Lowering this value too much means it may select more than you intend.
Press Generate. That's it!
|as a delimiter. For example, if you enter
a face|a tree|a flowerthen clipseg will process these three items independently and stack the resulting submasks into one final mask. This will likely yield a better result than had you searched for
a face and a tree and a flower.
Mask Paddingoption to increase the boundaries of your selection. For example, if you enter
a red shirtas your prompt but find that it's not quite selecting the whole shirt, and
Mask Precisionisn't helping, then you can use padding to address the issue.