Reply to thread

Message

<blockquote data-quote="Blue" data-source="post: 8970467" data-attributes="member: 20564">I've got a Stable Diffusion version installed on my PC, but I've barely got enough VRAM (6GB) to do things at a reasonable size.&nbsp; One thing I have been playing around with is a lot of different models which can get you rather drastically different results.One of the discussion points that comes up every once in a while is how &quot;horny&quot; the model is.&nbsp; Because one of the reason people run it locally is to have an implementation that disables the SFW filters.&nbsp; So lots of the models out there are trained including NSFW images.&nbsp; And the more that models will include risque elements where the prompt is silent is how &quot;horny&quot; that model is.&nbsp; I often have &quot;sfw&quot; as one of my prompts and &quot;nsfw&quot; in my negative prompt.Another big point about the models is tagging.&nbsp; There is a large anime-style repository called Danbooru that was used to train a model called Waifu Diffusion, which then has gotten mixed into many other ones.&nbsp; But Danbooru has in-depth tagging done by people for years and years and are well documented, so it kept those tags.&nbsp; But those are different tags than used elsewhere, so prompting can get odd - especially when mixing Danbooru and non-Danbooru models.</blockquote>

[QUOTE="Blue, post: 8970467, member: 20564"] I've got a Stable Diffusion version installed on my PC, but I've barely got enough VRAM (6GB) to do things at a reasonable size. One thing I have been playing around with is a lot of different models which can get you rather drastically different results. One of the discussion points that comes up every once in a while is how "horny" the model is. Because one of the reason people run it locally is to have an implementation that disables the SFW filters. So lots of the models out there are trained including NSFW images. And the more that models will include risque elements where the prompt is silent is how "horny" that model is. I often have "sfw" as one of my prompts and "nsfw" in my negative prompt. Another big point about the models is tagging. There is a large anime-style repository called Danbooru that was used to train a model called Waifu Diffusion, which then has gotten mixed into many other ones. But Danbooru has in-depth tagging done by people for years and years and are well documented, so it kept those tags. But those are different tags than used elsewhere, so prompting can get odd - especially when mixing Danbooru and non-Danbooru models. [/QUOTE]

Verification