自動翻訳は、かなり変ですが修正する能力が私にはありません。The automatic translation is pretty weird and I don't have the ability to fix it.記事の方が解り易いかもしれません。An article might be easier to understand.<a target="_blank" rel="ugc" href="https://civarchive.com/articles/1518/lorai-tried-learning-the-background-of-lora">Civitai | LORAの背景学習させてみた、I tried learning the background of LORA</a>v1.5背景学習比率、ノイズ1：白背景500：灰色20。白背景の学習タグにwhite background,simple background,を2回記入。学習モデルの変更。Background learning ratio, noise 1: white background 500: gray 20. Write white background, simple background, twice on the learning tag on the white background. Change learning model.以下、変更過程ノイズ背景を使用せず、学習タグにwhite background,simple background,を2回記載（背景そのものをタグ化して呼び出さない限り影響力を下げられないかテスト）一定の効果があり、背景描画が若干上昇した気がする。Below is the change processWrite white background,simple background, twice in the learning tag without using noise background(Test whether the influence can be lowered unless the background itself is tagged and called)There is a certain effect, and I feel that the background drawing has risen slightly.学習モデルの変更、背景をより詳細に描けるモデルへ変更白背景学習だが、一定の効果がある気がする。Change the learning model, change to a model that can draw the background in more detail It's a white background learning, but I feel like it's going to be a certain effect.背景学習比率を、ノイズ1：白背景500：灰色20へ変更。背景の不安定さはかなり低下、ノイズ背景は学習させすぎるとパターン学習するらしい。Change the background learning ratio to 1 noise: 500 white background: 20 gray. The instability of the background is considerably reduced, and the noise background seems to learn patterns if it is made to learn too much.ｖ1.5はこの3項目を全て適用してみた。教師画像が少なく、多様な背景を学習できない場合は、この方法も有りなのではないだろうか。v1.5 tried to apply all these three items. If there are few teacher images and it is not possible to learn various backgrounds, this method may also be possible.v1背景画像修正テスト今まで、白背景9、灰色背景1、の比率で学習させてきましたが、良くも悪くも背景画像の単調さがありました。今回、背景画像の比率を白背景3.5、灰色背景0.5、ノイズ背景6、の比率で学習させて背景の単調さが解消されるかテストしました。テストなので学習キャラはhaman様のみ。背景が、かなり不安定です。望む背景は得られないかもしれません。（ノイズそのまま出たりします)推奨設定weight0.5Hires. fix　２倍ADetailerを使用しface_yolov8nで顔修正多くは試していません。Background image correction testUp until now, I've trained with a ratio of 9 white backgrounds and 1 gray background, but for better or worse the background image was monotonous.This time, we tested whether the monotony of the background is eliminated by learning the ratio of the background image with the ratio of white background 3.5, gray background 0.5, and noise background 6.Since it's a test, the learning character is only haman. The background is rather unstable. You may not get the background you want. (Sometimes noise comes out as it is) Recommended settingsweight0.5Hires. fix x2,Face correction with face_yolov8n using ADetailerI haven't tried many.

haman karn LORA(GUNDAM ZZ)background test

背景ノイズテストbackground noise test

Files with SHA256: 8e3041789dcd572f596bf2b35bd2fbe4722b7a752db4a39146c9e535a8c45902