And thank you for commenting! If you can be bothered, please return with a comment on what you think of it once you've given it a good spin. Even if your impression is negative. Any feedback is helpful, because I use pretty unorthodox prompting techniques myself, so I really don't have much of an idea of how it performs for other people who might use a more standard approach, with "bad hands" and whatnot in the negative (I never do this myself).

@lilquin I'm happy to see I'm not the only one that feels this way! SD 1.5 is an artist in the true sense of the word. It is pretty stupid and doesn't take detailed instructions all that well, but it makes up for it in spades with creativity and a subtle subconscious understanding of the human condition. SD can make you feel something, where other newer and "better" models just produce pictures that look like bland corporate approved PR material. SD 1.5 is super unsafe and can and will hurt someone's feelings at times.

But that's just what real artists do.

I don't really have a version without the LoRA, since the final version was merged together from several work-in-progress versions that all had it merged in already. But all the ingredients had it merged at ratio 1.0 so you can easily remove it by simply merging it in again at a ratio -1.0! For instance using the A1111 addon Super Merger. I don't think this would make you any happier though. This particular distillation LoRA does more than just speed up generations, they also baked in some form of "aesthetic tuning", and all my testing and tuning was done with it in place, so I think it's an integral part of the model, and it would most likely get worse without it. Feel free to try tho!

About 20k, mostly people. By the way, I wasn't entirely happy with how v1 turned out, so I'm hard at work with v2. It's basically going to be made from the same data set, with a few additions, but mostly just tuned differently. I'm already getting vastly improved results, so I'm feeling very confident about it! v1 produces overly saturated images on some seeds, this issue will be fixed, skin detail will also be improved, and anatomy will be improved as well. At my current rate of work I expect v2 to be done in two weeks (or thereabouts), so do check back later, I think you'll be pleasantly surprised!

@Olbanets As I said to the other guy who asked the same question, no, not really! I understand you are skeptical about acceleration LoRAs, I was too. But this particular one is different from all the others. It is much more subtle to begin with. Most acceleration LoRAs try to make pictures at ridiculously low step counts like 1 or 4, but this one targets 12 steps, which is perfectly reasonable and gets good results. It also has some form of aesthetic tuning built in, which gives a certain attractive "sheen" to the images that I like, so I quickly found that it not only speeds up generation, it improves the model as well. And even if it is designed for 12 steps you can still use both less or more if you so prefer, so it does no harm. Since I liked the results of adding this LoRA so much I decided to do all tuning on the model with it baked in, at a ratio 1.0 . It can therefore be removed by adding it again at ratio -1.0 , but then you get a model that is knocked off the carefully tuned balance point I spent so much effort to achieve. I think the only reason to do so would be if you wish to use it for merges, that's why I also publish a F32 version.

EDIT:

OK, I tried, but civitai renames files so it's impossible to tell which is which, so I took them down again.