Avoiding Getting Blurred NSFW Generations - TravelingRobot/NAI_Community_Research GitHub Wiki

So if you somehow don't know, NovelAI currently offers a little preview of their upcoming image gen on their discord server. You can hop on to #image-bot and generate your own amazing AI art there!

However, unlike in the upcoming image generation within NovelAI, generation of NSFW pictures on the discord is not allowed. NSFW images will be automatically detected and blurred. Occasionally, NSFW generations can slip through the filter, though. In that case, just dm the modmail and a mod will delete it for you.
Users are expected to try to avoid using prompts that keep generating NSFW. Continuing to use the prompts prone to generate NSFW unadjusted will be interpreted as intentional generation of NSFW content. Intentional generation of NSFW content will result in restriction.

Now, it can very well happen that you realize that your seemingly innocuous prompt is somehow prone to generate a NSFW. What is going on?

Disclaimer: This is based on personal observation and opinion. The following does not necessarily reflect the opinion of the NAI mod team.

Why Do My Generations Keep Getting Blurred / Generate NSFW?

Think of Stable Diffusion / the image bot as a very talented artist… that is also a horny teenager. Any mention of terms that are even remotely related to nudity, and it might think of nothing else (especially if you keep adding up such keywords). Don't blame the poor AI, blame us humans for our training data…

There are some specific keywords I see frequently used that can lead to problems. Not all of these necessarily lead to NSFW gens. Sometimes it is just the effect of several keywords adding up. The main use of this list is to know what to look out for and to know which keywords to try taking or swapping out if a prompt proves prone for NSFW generation. Obviously, this will not be an exhaustive list, but might give you an idea what to look out for:

The AI Does Not Understand "No"

  • sfw and safe for work. Due to how the underlying language model works, these keywords actually make NSFW more likely! Language models can have trouble distinguishing positives and negatives. For the model, sfw looks very similar to nsfw so they are interpreted similarly. See [experimental methods with prompt mixing] below if you want to try to dissuade NSFW content.

Don't Make It think Of Naked Skin

  • Unsurprisingly, keywords like the following rapidly add up to NSFW gens: breasts, cleavage, exposed (even if followed just by something harmless like belly, shoulders, etc), etc
  • Keywords describing body shapes can steer towards NSFW, curvy, thicc, skinny, etc.
  • <color> skin (dark skin, pale skin, etc): skin is associated with NSFW. <color>-skinned seems less of a problem (fair-skinned, dark-skinned, etc).
  • Some colors often used to describe skin colors are heavily associated with NSFW as well: pale and ebony for example. Use fair or instead of pale and dark or black instead of ebony.
  • Clothing that imply bare skin can add up to other keywords to result in NSFW. Clothing keywords like stockings, underwear, bikini, swimwear, etc.

Don't Make It A Bloodbath

⚠️ **GitHub.com Fallback** ⚠️