I'm curious, do you know if there's a tutorial on how to make textual inversions/loras and whatnot? I've wanted to tinker with it but I'm not really sure how to get started, also not sure how far 6gigs of vram gets me.

@MiniMage For me, I've begun my learning attempt and they have gone alright so far. 6GB might be a bit rough. I have 8 and I have to turn some things down... but you don't (or at least I don't) have to trade quality, you just have to trade more time and do things a bit slower.

The most helpful thing for me was this video by Aitrepreneur: https://youtu.be/2ityl_dNRNw

But I took what I learned from that but also adapted some other things, mainly from reddit.

I've made about 5 of them now myself. Here to me is the 3 most important things (directly from my notes):
1. TURN OFF CROSS ATTENTION OPTIMIZATION WHILE TRAINING IN SETTINGS (every single training I did before doing this looked NOTHING like what I was going for). According to another source, this may be an error setting in Auto1111 and may get fixed later.
2. With that setting off though, with my 8GB VRAM I was basically forced to use only 1 a batch size and gradient accumulation of 1. Every time I tried to do larger it would give me an error. But that's alright, it can get you a very accurate training (multiple other reddit posts are what reinforced this for me).
3. Training rate is really important, this is ultimately (I think) what will determine how accurate your TI comes out as well as how versatile it is.
*Bonus is to train on 1.5 because that's what most models use as a base, so it will ensure the most compatibility with other models. Oh, and turn off VAE while training (I think. Not 100% on this part).

@Balthazar99 I've seen the video you linked, but that one come with two hurdles. I need to cut everything down to the 1:1 scale, I have a vast array of pictures, but the majority of them are not 1:1. The second is that I need to tag each image properly, which becomes rather tedious as well as time consuming. There's the auto tagging feature, but it also auto scales the images to 1:1. I have a vast collection of for example fantasy maps and whatnot, and if I made that into a lora that allow you generate maps for example.

Since I'm writing an english science fiction/fantasy book, I got a lot of reference material. But if I am able to cut down filesize into properly trained loras, I imagine it would be useful for myself, as I would be able to cut down filesize on my harddrive, and it might help others.

There's apparently a lot of different kinds of extensions and ways to make for example loras. I heard of something called Kohya, is that one good? https://github.com/bmaltais/kohya_ss

There was another one as well that I don't remember the name of though. I'm not sure which is best. Do you have a preference for any of them, or do you stick exclusively to texual inversions with 1:1 pictures? So you think I'll be fine with 6gigs of vram then?

@MiniMage I don't know. I'm not an expert by any means. I just got into AI stuff like 2 months ago. So far I tend to prefer Textual Inversions because I've seen a lot of really terrible LoRAs that just don't seem all that accurate. LoRA seems all the rage right now, there was a lot less a month or so ago but now it's an explosion. And not an explosion of quality either IMO.

Additionally, at least with the Automatic1111 interface, it seems like Textual Inversions play nicer with other stuff, whereas LoRAs have their quirks with how they not only work with other LoRAs but also VAEs sometimes. LoRAs seem powerful for achieving a certain affect, but maybe less versatile overall. IDK, I'm still learning man.

As for the 1:1 thing, the website they talk about birme or whatever it is, it's easy enough to use, and as for the descriptions, well you get what you put into stuff. If you want to put in the effort, the potential is there. Keep at it man, or keep researching or whatever.