I am finally getting close, after hours of research and work, I am finally getting to the point thanks to Loras where I can create custom training modules of Dante from my drawings and then integrate them with existing furry AI models to produce the cutes! Okay, so I often throw out dozens and dozens of images fine tuning and after a whole night, this was my final result... and it still forget the ear dashes, but you can see it THOUGHT of adding them (those shadows are a bit odd in the ears). I might even go back and add those in. It wouldn't be hard.
Either way O_O. My 3070 is surprisingly talented. Not to say that I didn't spend a lot of weekends getting to this point, and I still have far more training and probably future improved models before I start getting some extra real rewards, but... this isn't bad. I might actually be able to put this BACK into the training model and add it as an extra data point - which has insane implications (the AI, with my guidance could effectively TEACH itself what Dante looks like).
That said, it's been a ton of fun and OMG. I LOVE that background and those clothes. I wanna go there O_O. It's so pretty! Don't you wanna read in that library?!
---------------------------------
As in TRYING to keep to the rules, here were my specs.
Positive Prompts: male, boy, furry, fennec, (DanteAffinityXD), head shot, exploring inside a fantasy sacred library, standing, boys adventurers outfit, fantasy clothing, oversized clothing, anthro, fluffy, fennec ears, ornate bookshelves, dusk, twilight, stained glass, inside fantasy library, cathedral, ancient books, study tables, gold leaf engravings, tile floor, (highres), epic, ( Adorable), (Detailed face), Great Composition, (Extremely detailed CG 8k wallpaper), (((highly detailed background))) <lora:DanteAffinityXD:0.49>
Negative Prompts: adult, muscles, monochrome, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, lowres, bad anatomy, bad hands, text, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, (((missing arms))),(((missing legs))), (((extra arms))),(((extra legs)))
Other details: Euler a, 150 sampling steps, 512x512, CFG 7.5
It requires 40 or so pictures, but I split them into about 71 pics, so many were high resolution so the AI could focus on different images, but I also cropped each one into four sizes (512x512, 512x768, 768x512 and 768x768) hand cropping each of my pictures to grab information, with a bit of focus on the face for closeups for some. They all had to use the same color scheme and style - so that limited me to drawings I did myself, which meant I couldn't use a huge chunk of my gallery and even then I had to throw out some pictures because I realized that it was learning things wrong from say "Dante grabbed his tail, so now the AI believes tails grow out of hands O_O" among other situations where the AI was learning things wrong that I could kind of gather from the results, and then I had to go back to my training set and ask "Why?!"
More pictures would obviously help, and I also realized I needed a mix of backgrounds, because when I used all plain backgrounds, it learned Dante, but then it forget how to draw a background, and when I added all backgrounds, it learned the styling on the background and couldn't draw Dante. So I have pics where I turn the background on and the same picture but with it turned off, and then put in a solid color like pink or white which I described to the AI. I made the pics, though, so I had access to the original layered files to do this :). The advantages of being me >:D.
Describing each file in the text file is the hardest, and I eventually came down to a strategy of focusing on Dante and then used the following pattern.
Notice the solo in there, all the pictures you have must include the character and only the character. So... more restrictions. I used my name as a tag inside the training data after describing the basics, for me, {species} is fennec and clothing is often short shorts, hoodie. Don't include colors, it only confuses the AI and it will decide your character MUST be pink. Posture is also hard, but think like a booru, the blipping from the boorus on Automatic 1111 help a ton with this as well as figuring out your backgrounds, but your life is easy if you can use simple background, green background, or whatever. The color matters here ironically.
But I opted for about 35 repeats on my training model and 4-7 epochs verse his 1. Also I train one image at a time verses several, which keeps it under my VRAM limits. Lots and lots of retries O_O, I have been trying on and off to get this to work since last November, when I was excited to see AI produce a blue fox after hours of training Dante XD. "Yes yes! And he has big ears! Come on! You can do it little robot!"
O_O I am now happy with the result, and I am super excited to see what will come next :D
Sorry. TL;DR: Lots and lots of work to make your own Lora but then, think of what you get! :D
It requires 40 or so pictures, but I split them into about 71 pics, so many were high resolution so
That's both exceedingly convoluted and exceedingly interesting at the same time XD It's really impressive that something like this is even possible, even though it still made some mistakes it's really well made, I wonder what other ideas you'll come up with using AI 🤔 ;p
That's both exceedingly convoluted and exceedingly interesting at the same time XD It's really impre
