Only a small percentage of Stable Diffusion’s dataset — about 2.9% — NSFW material, giving the model little to go on when it comes to explicit content. So the Unstable Diffusion admins recruited volunteers — mostly members of the Discord server — to create porn datasets for fine-tuning Stable Diffusion, the way you would give it more pictures of couches and chairs if you wanted to make a furniture generation AI.
While the idea of Hentai diffusion sounds promising as porn is one of the never-ending industries, it takes a while to train stable diffusion models for Hentai. Thanks to
It’s well-known that porn is a of new technologies, and has been ; so whatever your view may be on the growing corpus of NSFW material that’s emerging from the new Stable Diffusion communities, NSFW ‘user enthusiasm’ seems as likely as it ever was to come up with new and innovative techniques for driving the technology forward into more mainstream applications.
Additionally, there is nothing to stop NSFW Stable Diffusion communities from developing exclusively pornographic video datasets that would be far more efficient at generating this kind of material after training. They are already doing so for static porn datasets:
In the meantime, a new wave of SFW and NSFW implementations and APIs of Stable Diffusion are likely to emerge in the next six months, many of which will seek to differentiate themselves by offering ‘fine-tuned’ models that have added domains (certainly not excluding porn, in some cases) that are otherwise under-represented or absent in the official sources used by Stability.ai – or which offer facile training or fine-tuning, so that users can develop their own ‘specialized’ models – for a price (likely a subscription-based price).
