Serg
Additionally, nearly all of these web-based offerings refuse to output the NSFW content (much of which may relate to non-porn subjects of general interest, such as ‘war’) which distinguishes Stable Diffusion from the bowdlerized services of OpenAI’s DALL-E 2.
Nonetheless, experiments are currently taking place within the various Stable Diffusion Discords that use much higher numbers of training images, and it remains to be seen how productive the method might prove. Again, the technique requires a great deal of VRAM, time, and patience.
In the example below, which features almost no movement at all from the (real) blonde yoga instructor on the left, Stable Diffusion still has difficulty maintaining a consistent face, because the three images being transformed as ‘key frames’ are not completely identical, even though they all share the same numeric seed.
Nonetheless, the way that aesthetic groupings bind together in these interfaces can teach the end user a lot about the logic (or, arguably, the ‘personality’) of Stable Diffusion, and prove an aide to better image production.
