Many thanks for jibunsagasinotabi san help, although my version is still not good enough , it is can't mproved without his help . Thanks for the help.
His version should be more better, Thanks for support this model.
https://civarchive.com/models/179815?modelVersionId=194135
For IL version
Nora Polyansky, Long white hair, hairclip , red lips,grey eyes
Pilotsuit,,gloves,pants,shoes,
color lens pilot helmet <---- may need around 1.2 streath)
purple military uniform,gauntlet,pants,shoes
For pony version
Upscaled, clipped photo and change some keyword
Bodysuit --> military uniform
For sd1.0 version:
Necessary Trigger word Nora Polyansky, Long white hair, hairclip , red lips
Optional Trigger word: Pilotsuit/ bodysuit/spacehelmet(strongly recommand use at close up picture and with a higher strength)
Smirk, angry
Extremely angry, Extremely smirk: a enhance version of angry/smirk , similar to 顔芸(kaogei)
FOR NORMAL BODYSUIT VERSION (former)
28/10 update : i found that you can just Press Nora to trigger ORZ, sorry for any inconvenience caused
Necessary Trigger word: Nora , 25 years old , White hair, red lipstick, bodysuit, dark eyes, white hair clip
or
you can type the sentence i used in training model :
Nora , 25 years old, White hair and dark eyes, wearing red lipstick and white hair clip , and she wearing bodysuit that cover her body exclude head.
FOR PILOT SUIT VERSION
refer to trigger work coloum
optional Trigger word:
Angry face, Smile face, smirk
Side view
Back view + looking back
Todo:
Refine both of the version ( some detail is lost in this version ORZ, also pilot version are not completely finished )
Description
FAQ
Comments (12)
I think your attempt to make the dataset public is great.
I will also try creating one using this dataset.
データセットを公開しているあなたの試みは素晴らしいと思います。
このデータセットを使用して、私も、作成してみます。
thanks for your comment, TBH i am still learning how to create a model (especially learning how to describe a photo LOL, thats made me very confused haha )
I will be glad to see someone create a better version. Please feel free to use, just please left a credit once if you publish a version ^_^
@Buddy17
I think the problem with creating it is in the teacher image.
The grain is rough compared to the number of pixels, and the dots are scattered. Please enlarge the teacher image and check it. You will see that it is not uniform.
I think it would be a good idea to clean it using "R-ESRGAN 4x+ Anime6B" etc. in "Extras" of "Stable Diffusion".
作成についての問題点は、教師画像に有ると思います。
画素数に対して粒子が荒く、ドットが散らばっています、教師画像を拡大して確認して下さい、均一では無い事が判ると思います。
"Stable Diffusion"の"Extras"で"R-ESRGAN 4x+ Anime6B"等を使用してクリーニングすると良いと思います。
I think it is better to erase the black screen outside the image in the teacher image. "Stable Diffusion" learns in squares, so if there are unnecessary parts, the image to be learned will become smaller.
教師画像にある、画像外の黒い画面は、消した方が良いと思います、"Stable Diffusion"は正方形で学習するので不要な部分が有ると、学習する画像が小さくなります。
https://civitai.com/models/179815/nora-polyansky-macross-zero-lora
It is necessary to separate the folders and change the number of learning times depending on the quality of the images.
If you want to implement special costumes such as pilot suits, you will also need to separate the triggers by dividing them into folders.
画像の質でフォルダーを分けて学習回数を変化させる必要が有ります。
パイロットスーツ等、特殊な衣装を実装したい場合もフォルダー分けしてトリガーを分散する必要が有ると思います。
thank for the comment and taught !
it really help for me refine the model! ^^
i will try to update the model with through guideline you say later
@Buddy17
After trying it out, I found that it is more effective to divide the helmet into folders instead of the pilot suit.If you don't attach a trigger such as a "space helmet", it becomes a normal helmet.
試してみた結果、パイロットスーツではなく、ヘルメットをフォルダー分けした方が効果が高いです、"宇宙ヘルメット"等のトリガーを付けないと普通のヘルメットになってしまいます。
@Buddy17
I have uploaded the teacher image used for v1.2.
I added one AI image and a few images borrowed from the internet, and deleted a few images.
I increased the number of steps for close-up images of the face, and decreased others.
The close-up of the helmet has been separated into a "space helmet" trigger.
v1.2に使用した教師画像をアップロードしました。
1枚のAI画像と数枚のインターネットから借りた画像を追加し、数枚の画像を削除してあります。
顔のアップ画像のステップ数大きくし、他は少なくしました。
ヘルメットのアップを"宇宙ヘルメット"でトリガーを分けました。
Thanks for guideline ,i am now more understanding how to train a model@jibunsagasinotabi
@Buddy17
I am creating LORA through trial and error.
I'm glad that the number of anime character LORA researchers is increasing.
試行錯誤しながらLORAを作成しています。
アニメキャラクターのLORA研究者が増えるのは嬉しいです。
@jibunsagasinotabi
Hi , i am Buddy, i just release a update a new version and provide a dataset with more Nora wear Pilotsuit scene. Although my version is not good enough , your help are really improved my model and make me understand more how to make a good model. Once again, let me thanks for your help here! <3
@Buddy17
I think it has improved a lot!
This technology can be applied to other things, and there is a possibility that it will become even better by replacing the teacher image with a higher quality one.
Even if the image is of high quality, it is important to enlarge it and check that there are no grainy grains.These grains can cause erroneous learning, so be sure to remove them.
There are also places for more advanced discussions other than teacher images.
-------------------------------
とても良くなったと思います！
この技術は、他に応用する事が出来ますし、教師画像を高画質な物に入れ替えることでさらに良くなる可能性が有ります。
高画質な物でも、拡大して粒子が荒くいないか確認する事が重要です、この粒子は誤学習の原因になるので取り除きましょう。
教師画像の事以外で、より発展した論議をしている場も有ります。
-------------------------------
@jibunsagasinotabi
Thanks for the reply, I will try to learn more from it (◍•ᴗ•◍)