This is SD1.5 version.
SDXL version：
【SDXL】Mai Shiranui 不知火舞 doa5lr LoRA ◤Louiswolf_style◢
Pony XL version：
【Pony XL】Mai Shiranui 不知火舞 DOA5LR LoRA ◤Louiswolf_style◢
這是我第一次發布這個角色的LoRA，我很喜歡這個角色，希望大家會喜歡，在這邊分享與各位交流。
This is the first time I've released LoRA of this character.
I like this character very much.
I hope everyone will like it.
I would like to share it with you here.
If there are images using this LoRA you create, I hope you can post images and comment with everyone, thank you.
一.我使用的設定 (My setting)
Sampling method：DPM++ 2M Karras
Sampling steps：20
CFG Scale：7
二.關於LoRA (About LoRA) (New version V 1.6.1)
1. Description:
This LoRA can images Mai Shiranui (Ver DOA5LR) images.
2.My way to use prompt:
Trigger Words :maishiranuidoa5lr
This LoRA can paint three hairstyles for this character.
prompt:
(1) ponytail(default hairstyle)
maishiranuidoa5lr, 1girl, solo, long hair, ponytail, high ponytail, hair ribbon...... <lora:maishiranuidoa5lr:0.6>(0.2~0.6, Decide by yourself.)......
(2) long hair, loose hair
maishiranuidoa5lr, 1girl, solo, long hair, loose hair...... <lora:maishiranuidoa5lr:0.6>(0.2~0.6, Decide by yourself.)......
(3) short hair
maishiranuidoa5lr, 1girl, solo, short hair...... <lora:maishiranuidoa5lr:0.6>(0.2~0.6, Decide by yourself.)......
Negative prompt:
In principle, none. Decide by yourself.
Personal preference prompts:
masterpiece, best quality, Full-HD, HD, 8K, HDR, highres, highly detailed, RAW photo, dreamlike, (Magical Photo:1.3) of (Realistic:1.3), (Controllable Photo:1.3) of (Ultrarealistic:1.3)
3.LoRA權重 (LoRA Weight) (New version V 1.6.1)
LoRA Weight:0.2~0.7 more ideal.
4.使用ADetailer (Use Adetailer) (New version V 1.6.1)
It is highly recommended to use ADetailer to inpaint the face to achieve the ideal effect.
ADetailer model：face_yolov8n.pt
ADetailer prompt：detailed face, beautiful face, perfect face <lora:maishiranuidoa5lr:0.7> (recommend 0.4~0.7, I use 0.7, decide by yourself)
ADetailer Inpaint denoising strength：0.4
5.個人使用的Stable Diffusion checkpoint (The Stable Diffusion checkpoint I use)
Some checkpoints are not highly suitable for this LoRA.
I personally recommend using it with AniMerge. This checkpoint is suitable for this LoRA and it can generate the ideal face and hairstyle.
You can also try other checkpoint to paint the style you want.
Happy painting and don’t hesitate to share. Thank you. :)
6. Important
My way paint Mai Shiranui with this LoRA
(1) Use LoRA and prompt to generate picture.
(2) Put picture you create in img2img.
(3) Use img2img (no prompt) to generate new picture.
Let the picture you create more naturally without rigidity.
(Denoising strength:0.3~0.8 {decide by yourself} )
(4) Put picture you generate in img2img inpaint.
inpaint face and what you need.
face prompt I use:
detailed face, beautiful face, perfect face, smile <lora:maishiranuidoa5lr:0.4>
(Denoising strength:0.3~0.5 {decide by yourself} )
Let face more naturally without rigidity.
(5) Finish.~~
7.Tips (New version V 1.6.1)
You can use Hires. fix to upscale pictures you create.
Make pictures more clarity and more like DOA5LR style.
Setting:
Upscaler:4x-UltraSharp
Denoising strength:0.35
Have fun.~~
Description
1. Update note:
Let hair ribbon more like Mai's hair ribbon.
prompt:
maishiranuidoa5lr, 1girl, solo, long hair, ponytail, high ponytail, (hair ribbon:1.15), upper body, medium shot, smile, looking at viewer, outdoors, day, sky <lora:maishiranuidoa5lr:0.65>......
Negative prompt:
poorly drawn ponytail, poorly drawn hair
2. Important:
My way paint Mai Shiranui with this LoRA
(1) Use LoRA and prompt to generate picture.
(2) Put picture you create in img2img.
(3) Use img2img (no prompt) to generate new picture.
Let the picture you create more naturally without rigidity.
(Denoising strength:0.3~0.8 {decide by yourself} )
(4) Put picture you generate in img2img inpaint.Inpaint face and what you need.
face prompt I use:
detailed face, beautiful face, perfect face, smile <lora:maishiranuidoa5lr:0.4>
(Denoising strength:0.3~0.5 {decide by yourself} )
Let face more naturally without rigidity.
(5) Finish.~~
FAQ
Comments (9)
In my opinion there is no need for the entire procedure described to create the image. The lora is created with a net that is too low and with too few epochs and too many steps. For a possible new version, try setting the net to 64 and making sure not to exceed 5000 steps while keeping the epochs above 10. This way you will have a more flexible lora with more information inside it. Good work.
Thank you for your substantial opinions.I just think that use many trigger and steps make LoRA what I want.
Consequently, I can just use 5000 steps (not to exceed 5000 steps), epochs above 10 and a lot of pictures to make new version LoRA?
@Louiswolf It depends on what you want them to learn. Basically, the more details you put in, the more flexible the lora is. The trigger words are usually just the first word of the caption. If you use the "shuffle" option and some sentences are repeated they will become fixed in the lora. A single trigger word is enough to focus a character. The size of the net serves to be able to insert all these details into the lora. If the lora is small, you will have less detail and useful information for prompts.
Could you give me a few pointers that "net" is means "Network Alpha"??Thank you.
@Louiswolf Try this settings : unetLR 0.0005, clipskip 2, keeptokens 1 (first word is a a trigger word), network alpha 64, shuffle caption on, networkdim 32 or 1 (try for better results), textencoderLR 0.00005, optimizer adamW8bit or AdamW (try). Now if you have 50 pics, try 2 repeat and 20 epoch. If you have 50-100 pics try 1 repeat and 18 epoch. If you have 100+ pics try 1 repeat and 12-15 epoch. Create captions with W14 and add or remove unnecessary things but insert details on the backdrop and costume. For training, use images that also show the hands and other body parts, as well as images that include the face in detail. The problem with some checkpoints not being good with LORA depends on the information that is in your LORA. Since it is too small, there is little information and trained models are needed to reproduce it faithfully. If you can, train the lora with a good realistic model and you can also use it on other anime or semi-realistic models without problems.
@MyIA Thank you.I will try again~~Thank you for your substantial opinions.
@Louiswolf Good work ;)
@MyIA Thank you for your substantial opinions.I train a new LoRA. It's better then before so much:)
I am grateful for your suggestions. :)
@Louiswolf I'm happy. If you need help write me. Remember those settings are fine for most of the loras.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.