At the start I'm training only pixelated censoring. I would like to first fine tune dataset/training on understanding of the concept of censoring on any object/part of object with natural language.

Version comparison

• ALPHA2 is much more flexible than ALPHA1

• ALPHA2 is able to censore things that were not present in training dataset

• ALPHA2 has a bit of issues with hands and quality (jpeg artifacts)

• ALPHA2 is not censoring things that you do not specify as ALPHA1 did

• ALPHA2 does not censore somtimes entire objects (mostly on bigger ones)

How to use

• Trigger word censored

• It works pretty well with natural language prompts (prompts that seemed to work for me are adding to the end "<something> is/are censored.", before the thing you wanna censore add "censored <something>", or add right behind "<somthing> that is /are censored")

• Don't use this LoRA on clip or keep it on strength 1

Known issues

• Jpeg artifacts in certain parts of the image

• Hands are generated blurry or with wrong number of fingers

If you have some ideas for future types of censoring leave them in the comments.

Enjoy :)