Illustrious、NoobAIベースのモデルを使ってマージしました
バージョンごとに大きく変わっているかもなので
お好みのバージョンをお使いください
branchBはv13.0の時で使用したモデルの最新版で再作成したもの
バージョンにplusが付いているものはLoRAマージあり(v19.0以降)
※旧バージョンではLoRAマージで版数上げていましたが今後は19.0同様にする予定
(google translation)
There may be significant differences between versions, so please use the version you prefer.
branchB is a re-creation using the latest version of the model used in v13.0.
Versions with "plus" in the version number include LoRA merge (v19.0 and later).
*In previous versions, we increased the version number by merging with LoRA, but from now on we plan to do it the same way as version 19.0.
サンプルイメージの設定は以下。reforge使用
The sample image settings are as follows:
Sampler：DPM++ SDE
Schedule type：Automation(Karras)
[Normal]
Step：6～12
CFG Scale：4～6
[DMD2 ver.]
Step：4～
CFG Scale：1
Description
アニメ調のモデルを追加
少しプロンプトに素直になった・・・と思う
※LoRAはマージしていません
Added anime-style models
I think it's a bit more responsive to the prompts...
※Not merging LoRA
FAQ
Comments (7)
AIコスプレ用にずっと1.4を使っていたのですが、3.0良さげですね（スケージューラーが限られる感じがありますが）
I've loved the Kodora models since day one, however I don't know what happened between v3.0 and the previous version, but v3.0 consistantly generates blurry and blocky images. Hair especially no longer looks realistic and instead looks like painted plastic shells. Compared to pre-v3.0 images using the same prompt and settings (even in the sample images,) the older versions looked beautiful and all the v3.0 results look like poorly-compressed .jpg images. :(
Thank you. This may be due to the addition of anime-style models. I'm sorry, but please use the old version or the one you prefer.
※google translation
I'm not sure if it will meet your expectations, but I've also uploaded the version before I added the anime-style model. (v2.5)
※google translation
Thanks for the upload of 2.5. It does seem to suffer from the same issues as 3.0, though not to as such an extreme level. I've posted an image comparing the last four versions using the same prompts/settings/seed. You can see that skin textures, hair, and even lighting, seemed to peek with 2.1LH. Then something happened with 2.5 that got even worse with 3.0.
In that case, it's possible that this is due to the influence of other models that were added.
I can't promise that it will be improved in the next version, but I'll take your information into consideration.Thank you.
※google translation
v3.1 is looking good, doesn't seem to have the problems the last couple versions did. :)