Type | |
Stats | 262 |
Reviews | (11) |
Published | May 26, 2023 |
Base Model | |
Training | Steps: 2,000 Epochs: 10 |
Trigger Words | RemiliaScarlet2009 pixiv2009 |
Hash | AutoV2 1275D54BA9 |
The dataset for this one is 20 images from pixiv.net.
All of them are fanart of Remilia Scarlet from the Touhou Project game series.
Come to think of it, I should have thrown in her official art as well.
Well, anyway, all the dataset images are of Remilia, but since it's a LoCon, it can actually be pretty flexible and make any character look like they're from a 2009 fanart image.
It can work well combined with my HL2 Concept Art LoRA.
This model can be a bit picky; when I tested it on profantasy_v2, it would only work with a unet weight of 0.5 or lower, but on dreamshaper_4BakedVae, it would work at any weight.
It seems to work best on models that have at least a basic amount of training on Touhou-related images.
I created this model - and the HL2 concept art one! - because I love the various art styles of the 2000s, and even before the advent of NovelAI I was lamenting at how art had become so homogenous in style and flatly lit.
You even see it in video games these days. A wall texture is no longer a detailed surface, it's just a solid colour with some faint brush strokes. It may be easier to pick out players this way, but it's also stylistically bankrupt while claiming to be the opposite. Just criminal.