While training dreambooth for cartoon images, sometimes very simple, pleasing looking cartoon images get flagged by the safety checker model as NSFW. Disabling it is not an option however, since with bad keywords, the model does have potential to generate NSFW images when made available publicly.
What approach could be taken in this scenario?
- Is there any getting-started guide / documentation to fine-tune safety checker model?
- Are there different flavors/severity of safety checkers available to choose from as plug-n-play to see which works best?
- Can any preprocessing be done to input images of dreambooth to assist generated images pass through safety check?
Thanks in advance!