MISSILE228 Style Painting版本v2.0 (ID: 34208)

Copyright

Guidelines for use

This model and training set are for educational use only, not for commercial use.

When you use this model to generate images or use the training set to generate models to generate images, you should indicate the source of the model and the pid of the artist.

Painter pixiv user id 429077.

https://www.pixiv.net/users/429077

Warning again: You should not use this model without acknowledging the artist id, nor should you use this training set for other purposes.

About dateset watermaker

For learning purposes, I removed the watermark on some of the author's images. You should not use the watermarked images for any purpose other than learning. You must promise that you use these pictures only for the love of the picture creators.

Copyright statement of the author of the image

イラストの転載・使用に関するレギュレーションです▼

作品に関しては、「他作品等への提供物ではないもの・オリジナル作品」は転載（SNSや掲示板への掲載等）を許可します（趣味イラストなど）（常識的な範囲内での転載行為は許可してるので、メッセージでのご連絡などは無くても大丈夫です）　

ただし「二次創作作品」の転載や、作品の商用利用、に関しては連絡をください

創作（動画等）の使用に関しては、作品の設定を壊さない程度であれば使用を許可しております（音楽動画のサムネなど）

ただし、作品に対し設定を加えた上での使用（シチュエーションボイス動画等）に関しては、設定の変化や勝手な付加が起こって欲しくないため、無断での使用は禁止しております　もしどうしてもそのような使用をしたい場合、使いたい絵と、それに対し加えたい設定やシチュエーションをセットでご相談いただければ検討いたします

個人的な範囲での使用（PCやスマホの壁紙　ＳＮＳのアイコンやヘッダー画像など）に関しては、改変しなければ自由に許可しております

動画などの創作や個人的な所有物に対する絵の使用は「転載」ではなく「作品の使用」に値するという認識です

Description

Model

recommand base model: counterfeitV25

lora recommand: about 0.7

The model is good at expressing light, and at this point it learns well from the original artist. But it does poorly on night scenes. If you use the promt of the picture example I posted, the random seed usually performs well. It may not perform well when you use more detailed descriptors.

The training of this model is done at a resolution of 1024*1024, I strongly recommend you use a resolution of around 768 for painting.

When the weight of lora increases, the picture will be destroyed, and the eyes of the characters will become ugly.

This is an overfitting model, and too many promts will also cause the screen to collapse.

I've been working on the 3090 for over a week and it still can't perfectly replicate the artist's beautiful style. I'm sorry but I did my best.

Trianing

I tried using Loha, but the model struggled to converge and the anatomy of the characters was very poorly broken. I have tried the following parameters (linear layer alpha = 1, convolutional layer alpha = 1):

linear layer dim = 128, convolutional layer dim = 32
linear layer dim = 64, convolutional layer dim = 32
linear layer alpha = 64, convolutional layer alpha = 16
linear layer alpha = 64, convolutional layer alpha = 8
linear layer alpha = 32, convolutional layer alpha = 16

I also tried using Locon, but I experimented with less parameter combinations.

Every training is tens of thousands of steps.

I think loha doesn't perform well in the training of painting styles that are not obvious.

This released version lora dim = 128.

I also tried dim=192 and I don't think there is a noticeable difference in their performance.

Regularization

When the regularization weight is 1, 0.5, 0.25, 0.125, it is difficult for the model to converge. In fact, 0.08 and 0.05 also converge slowly, but 0.05 still has an overfitting problem.

Dataset

image429077: 133 img & txt

image429077hasreg: each img in separate folder, good for regularization

reg: reg_all

You can make 133 images correspond to 2000 regularized images. It is also possible to have each image correspond to a separate regularized image, which is generated using the promt obtained by the corresponding inversion.

If you want to use the overall regularized image, please use the python file in my compressed package and change the source image directory and target directory.

Integrating helps to get a consistent style, training separately helps to get interesting results.