v2.0 trigger words:
original : mjr
append: vocaloid append
punkish: pnk ("no shorts"to remove shorts)
stylish energy: ste
strange dark: strd
white edge: wte
blue moon: blm
lightning stone: lts
trickster: tks
vivid bad squad: vbs
leo/need: lnd
more more jump: mmj
wonderlands x showtime: wxs
25-ji, nightcord de: njg
kgl: random clothing
==================
v1.0
Support costume and trigger words:
original custume: mjr, kagamine len
append: vocaloid append, kagamine len
punkish: pnk, kagamine len ("no shorts"to remove shorts)
stylish energy: ste, kagamine len
strange dark: strd, kagamine len
white edge: wte, kagamine len
blue moon: blm, kagamine len
lightning stone: lts, kagamine len
trickster: tks, kagamine len
vivid bad squad: vbs, kagamine len
leo/need: lnd, kagamine len
more more jump: mmj, kagamine len
wonderlands x showtime: wxs, kagamine len
25-ji, nightcord de: njg, kagamine len
In most cases, it works best with lora strength at 0.8~1.0 and trigger word strength at 1~1.2
Description
FAQ
Comments (6)
"In most cases, it works best with lora strength at 0.8~1.0 and trigger word strength at 1~1.2"
Just a note; this may be different if you want realistic output. Cuz in order to allow style changes(a cartoon/anime outfit to blend with a realistic style), you sometimes need to reduce the strength of the cartoon/anime lora.
PS: I like the fact that this lora is larger in size. Many people make the mistake of using the same lora rank/network_dim on multi-outfit loras as they use for single outfit character loras, despite the fact that many single outfit character loras still fail to get the outfit right.
Yeah, loras at first were all huge, in the 1gb range each, but then lora creators learned that they didn't need to use that large a lora to encapsulate their simple concept or character. but then, they never seemed to realize that they do need to use a larger lora for loras with multiple outfits or multiple characters, etc. It's like there was a disconnect between how much data it needs to be able to distinguish vs size. So a lot of them are like JPEGs that have been compressed down to 20% and are very lossy as a result.
@dita Yea, sad. Many people just don't go for the quality. Like, as a general rule. But even on single outfit/character loras, the smaller size does them injustice. Cuz on so many ive used, it's near impossible to reproduce the finer outfit details.
And that's not even touching on the datasets. I've downloaded some people's datasets, just outta curiosity, and 🤦♂️. Cuz A: it's obvious they used auto-captioning, even though many of these people have datasets of under 50 images. And didn't even bother to comb through and remove non-relevant captions.
Like, I've been through datasets where it was an all boy dataset, just one character, but it had more 1girl, than 1boy.. 🤦♂️
Then the images, not only are they not using good quality images(when better ones are available), but they burn the quality even further by training at 720-1024..
It doesn't help that all the tutorials tell people to train at those resolutions 'because that's what's in the bulk of the model training'. But then they don't realize that by doing that, the model rarely gets better even after dozens, or even hundreds of Loras have been merged, and the model's been fine-tuned, and even merged with other models that have had Loras merged in.
I seen this one guy put a training tutorial on YouTube for training an actual person lora, and he suggested to make images in which the subject is facing the viewer because "sdxl doesn't handle angles well" or something like that..
That's like telling someone not to teach their kid how to read, cuz they don't read well..
Maybe I'm just trippin balls here, but are you sure that you sure that you uploaded the right lora? under the metadata viewer the keywords are completely different..
It is the same lora that I used to generate showcase images. I renamed the Lora file before upload, but I was too lazy to regenerate all these images. I’m not sure if this is the problem that confused you.
@shounens Yeah, I suppose that one was MyBad.. It actually had nothing to do with the renaming. I am currently working on coding my own local metadata/safetensors viewer, using PySide6. And it functions great so far. Except, as I just discovered yesterday (after reading your response here), I will need to implement an alternative method for viewing Lycoris files; as it apparently does not display the tags correctly when using the same method as for Loras.