With the new nataili bridge we had lost the existing NSFW censorship model which was responsible for ensuring SFW workers who wanted to, did not generate NSFW images. It was also an important part of the censorlist implementation, where even NSFW workers could choose specific words which would activate the censor filter as they are concepts that never belong into pornographic prompts.
Today I’m happy the announce that with the help of hlky, the NSFW filter is back, so there’s no reason for anyone to stick to the old worker code anymore!
To add to that, I’ve also added an extra protection from what I call ‘Corrupt Prompts’ in the horde. That is, prompts which have a chance to generate completely unethical AI images. Now when the prompt detects the specific generation of words, the owner account and IP will be flagged and the request logged.
Nothing is perfect but I plan to keep tweaking the detection for corrupt prompts to ensure I prevent as much as possible from reaching the workers. I also hope that this extra logging will give pause to people who might think to use the horde for unethical purposes.
Unfortunately the safety_checker model responsible for the NSFW detection, takes a bit more of your VRAM to work, so if you find that after downloading the latest version of the bridge, your workers starts crashing, consider reducing your horde power, or unloading a secondary model, such as one of the dreambooths.