Word of advice: don't date anyone who is only staying with you because they cannot fly away.
I wrote out the prompt with variables in it, so I could change the variables or get them automatically filled in by randomly picking from lists. See https://github.com/adieyal/sd-dynamic-prompts#combinatorial-generation and the part about variables in the extension's help.
But the real thing that gives the images structure is the regional prompter extension, which interprets the ADDROW/ADDCOL stuff and makes the one thing be on one side of the image and the other thing be on the other side, by doing different parts of the image with different prompts in the same generation.
I think that's just how her nipple is?
Also more nostrils make horses go faster. This is well known.
Yeah, that was the point. Not having it come in until a step that doesn't happen should give identical output to not having it in the prompt at all, but doesn't seem to.
OK, I tried it via the maintained fork here, and it does something, but it doesn't really let you zero out a LoRa. Here is the model not understanding my prompt at all and drawing garbage, and then here it is when I increase the LoRa weight in the prompt from zero and it draws something different. In both cases I am using the extension to tell the LoRa not to come in until step 21 of my 20 step run. In both cases I told the extension to plot the LoRa weight it thinks it is using, and it was 0 at all steps. But clearly having the LoRa in there did something even when it was supposed to be at weight 0.
I think it's that you need to be able to throw parallel processing at a lot of RAM. If you want to do that on a PC you need a GPU that has a lot of RAM, which you can only really buy as part of a beefy GPU. You can't buy a midrange GPU and duct tape DIMMs onto it.
The Apple Silicon architecture has an OK GPU in it, but because of how it's integrated with the CPU, all the RAM in the system is GPU RAM. So Apple Silicon Macs can really punch above their weight for AI applications, because they can use a lot more RAM.
This was several different ones; it should be embedded in each image, but I think cyberrealistic_v31 worked well.
I tried it with LoRas; it works OK but the LoRa applies over the whole image, and I found that the LoRas seem to make the model less good at reading comprehension, which is what I am fighting with the regional prompter in the first place.
Really what I want is to use the LoRa for some generation steps but not use it for others, but I haven't worked out how to do that yet.
I mean, it kind of is? Not everybody has the same number of legs, so the model should turn out people with just the one as often as the real world does, unless the prompt tells it different.
Never seen anybody with quite so many knees though.