Supposedly 14 vectors 100k steps
Delutaya standing solo:
https://litter.catbox.moe/q7hout.pngYeah, mine doesnt fair that well either...
Twintail tests
https://litter.catbox.moe/ph15g7.pngSeem tough. What do you guys think again?
>>37976342From FAQ
https://rentry.org/sdgoldmine>What's the difference between embeds, hypernetworks, and dreambooths? What should I train?Anon:
I've tested a lot of the model modifications and here are my thoughts on them:
embeds: these are tiny files which find the best representation of whatever you're training them on in the base model. By far the most flexible option and will have very good results if the goal is to group or emphasize things the model already understands
hypernetworks: there are like instructions that slightly modify the result of the base model after each sampling step. They are quite powerful and work decently for everything I've tried (subjects, styles, compositions). The cons are they can't be easily combined like embeds. They are also harder to train because good parameters seem to vary wildly so a lot of experimentation is needed each time
dreambooth: modifies part of the model itself and is the only method which actually teaches it something new. Fast and accurate results but the weights for generating adjacent stuff will get trashed. These are gigantic and have the same cons as embeds
