Type | |
Stats | 108 138 56 |
Reviews | (22) |
Published | Jan 29, 2025 |
Base Model | |
Training | Steps: 30,000 Epochs: 5 |
Usage Tips | Clip Skip: 2 Strength: 1 |
Trigger Words | retroanimu retro_artstyle 1990s_\(style\) |
Hash | AutoV2 834993C1AB |
GOOD MORNING CIVITAI!
I come to you with another lora! I hold here, a product of frustration! There is no good retro style for Illustrious!
I hear you say, “What about this?” Well, dear straw-man chucklefuck, they don't count. The retro loras I've tried are either “pseudo” retro, or they're overbaked, overtuned, or just generally with little flexibility, or just come out as soup.
So, what have I done about it?
What I have here, is a combo of a dataset, assembled from actual period accurate screencaps.
We're looking at a dataset assembled from 150 images, after pruning, cleaning, and fixing from an initial 400+ images scraped between the good old ones of the 90s. Names not limited to:
Mobile Suit Gundam: Stardust Memory
Ghost in the Shell
Akira
Inuyasha
As you can see, we are looking at some fun picks.
After cleaning, I've exploited the on-site tagger to create captions. I was still dealing with some 200+ images, and then I filtered them again, using a python script, leveraging cleanvision. This left me with 150, which have been plugged in to Koyha, exploiting an adaptive optimizer to produce this fifth epoch I'm uploading after some 30k+ steps.
You would be surprised, but it's actually BETTER to have an undertrained lora for style, since seldom people use ONLY a style lora. Having it slightly undertrain allows better flexibility since less soup comes over from the style model. That in combo with a pretty wide rank/alpha allows the lora to expressively perform without having trouble muddying details up for any other loras injecting concepts or even other styles. What that means for you is you might need more strength not less with an undertrained LORA, but this also allows more freedom with expressing the style.
This on my 3090 took over 15 hours, so please donate some buzz if you can, or at least post your generations. It always makes me smile seeing people use what I made and having fun with it!
Now for the important things:
My suggested strength goes anywhere from 1.0-1.5, depending on your model, or your other Loras.
The trigger word primarily implemented is retroanimu however if your fancy desires more, you can add retro artstyle, 1980s (style), 1990s (style).
The dataset had ended up being more biased to the Gundam side of things, that means you can get some gundam adjacent features, like cockpits or flight suits or uniforms. Your mileage may vary, but that is inherently within the technology.