Hugging Face's logo

Aotsuyu
/
LoRA

Model card Files Community
LoRA / README.md
Aotsuyu's picture
Aotsuyu
Update README.md
6fcd440
preview code
|
raw
history blame
6.29 kB
metadata
license: creativeml-openrail-m

Arbitrary epochs I use the most.
Sometimes different variants of strength, which are usually different epochs.

The 'resized' folders contain the same loras resized with the resize lora script and sv_fro, they're similar in strength but much smaller.

The token for most LoRAs is either the LoRA name, it's in the LoRA metadata or it's not necessary at all.

Some previews included, but I didn't put much thought into them so they're not indicative of quality.

Artstyles

The symlinks folder contains symlinks to the same LoRA, using an older naming scheme.

  • WhiteMelancholy is a mix between yagamishiro, dElysium, closer to death's depressed waifus and xlimo.

    preview

  • bkubchibi is 9948. It really wants the chibi keyword to function properly.

    preview

  • crayoncirno is trained on cute little crayon cirno drawings.

    preview

  • dElysium is trained on disco elysium portraits.

    preview

  • darkKukiv1 is a really decent noise offset (not multi-res) LoRA, you can use it to generate darker pictures. It's quite strong.
    darkv7 is another noise offset LoRA, one that doesn't affect overally quality of the image much, but being weaker as a result.

    preview

  • deaver is 29999491.

    preview

  • faintpaint is a funky little LoRA mix between some pastelmix, whitemelancholy and diffLoRA to give it a more painty feel.

    preview

  • gotagotay is 30662116.

    preview

  • hkv2 is 5691902, hollow knight art. It can generate some interesting sceneries.

    preview

  • kasuyabaian is 78872596, trained on a dataset that hasn't been cleaned of text.

    preview

  • omao-locon is 29268484 though it's finicky.

    preview

  • pastelmix stronger and weaker follow the formula of the original pastel mix model (andite's) but merged into a LoRA with different weights.

    preview

  • rajawat is 32622636. I have no idea what this does.

    preview

  • reverinth is 66421798.

    preview

  • syokumurav4 is 5512911.

    preview

  • tekitoumidori is 50544926.

    preview

  • yagamishiro is 625564.

    preview

  • zunusama is trained on mostly his newer works.

    preview

Characters

  • SBKindred is Spirit Blossom Kindred.
  • bondrewd is Bondrewd from MiA.
  • erisallv2 is Eris Greyrat from Mushoku Tensei, with all of her outfits trained.
  • gwen is a League of Legends champion.
  • kndrdsks is Kindred, another LoL champion.
  • lain is Iwakura Lain. Lokr version included.
  • mdtsk is a 1dim Madotsuki from Yume Nikki.
  • ozen are Ozen from MiA.
  • Qiyana is yet another LoL champion.
  • Yazawa Nico is from Love Live.
  • bangkisks is the outfit/character from 暁Records's Rock 'n' Rock 'n' Beat PV.

Concepts

  • nnsks is a nose removal LoRA. Pendin retrain.
  • scuffedHorror is a noise offset LoRA trained on a horror dataset.
  • Hogwart Uniforms is a lora trained on the uniforms of Harry Potter's Hogwarts 4 houses.

    hogsks, hogwarts school uniform, black robe, gray vest, slytherin, green tie

diffLoRA

Magical little lora trained by overfitting on an image, then editing that image and training another lora and merging the loras with -1 and 1 weights.
This achieves a small lora that doesn't affect other stuff very much, mostly just providing the effect.

Most also yield interesting results at negative weights, and weights exceeding 1.

Due to how this method works, if the edited image is too different, the merge isn't that great and the original image might start showing.

WIP

Stuff that I needed to upload somewhere but I'm not sold on it being the best version.

archival

Mostly things that have a version that I find strictly better uploaded.