Great Minds...

[scruffynerf]

Just found this repo, and we basically did a lot of the same sort of digging into AceStep. I also did nodes to split conditionings apart, etc.
Haven't yet dug thru your code, but I will this week, to see if you did anything I didn't, and I'm in the midst of trying to make sure my code is 'ready for prime time' to share more widely (it's on github, if you want to look at it, https://github.com/scruffynerf/scromfyUI-AceStep)

Been doing quite a bit with acestep, so I enjoyed seeing that we seem to have approach it similarly. More after some code gazing...

[blepping]

Sorry for the lack of response (not that my responses to anything have been timely these days). I actually never got a notification for your issue. It's even that I just missed it, there's just nothing in the history.

I didn't have time to really look in depth at your stuff but it seems like while there's some overlap, also quite a bit of things unique to both repos. Most of the interesting stuff in here right now is related to the LLM sampling side. You can get some really interesting results using weird noise types for LLM sampling. Normal multinomial temperature sampling uses Gumbel noise but... what if we use Poisson noise or something even weirder? N-gram repeat penalties are also pretty nice for encouraging creativity.

One thing I think you're doing more of than me is manipulating the codes/conversion into the latent like space. That's something I've been meaning to work on but I haven't had a chance to yet.