QQ|Archiver|手机版|小黑屋|AiOMG ( 粤ICP备2022147020号-1 )

GMT+8, 2024-11-25 15:01 , Processed in 0.059709 second(s), 30 queries .

Powered by Discuz! X3.5 Licensed

© 2001-2024 Discuz! Team.

[LoRA] niji武士

27
0
0
0
simonlive LV9
正文
发布时间:2024-08-04
模型类型
模型类型: LoRA
基础模型: SD 1.5
最新版本: v1.0
触发词:
下载链接: https://pan.baidu.com/s/1hcdRDdlAIJj8Al8GNMqtAQ?pwd=eyhi
提取码: eyhi
源链接: https://civitai.com/models/89197?modelVersionId=94922
模型介绍:
一个niji风格的武士模型,欢迎使用,泛用性广,可搭各种2d底模使用,
由于拿剑的姿势很复杂,所以出好的手,几率超低,可搭ControlNet使用,效果会好,后续有空可能出一版没拿武器的版本,
图片都是简单背景训练。。要换复杂背景,可用低权重,然后背景加高权重试试,或者分层lora使用,

我目前测试的底模:dalcefoPainting_3rd,duchaitenniji_cuteV235,其它的自测哈,dalcefopainting_3rd链接 https://ko-fi.com/s/9dd78df21d
embeddings模型:EasyNegative,ng_deepnegative_v1_75t,badhandv4
VAE:vae-ft-mse-840000,animevae, 如果换底模颜色过重可换vae或者不用vae,
推荐权重:0.6-0.8 可用分层lora减少繁杂的特效。看个人喜欢。
Sampler:DPM++ SDE Karras,   
CFG scale:7

出全身要用大尺寸图,否则脸崩,我用的是512x768,高清修复放大2倍1024x1536,或更大,不要同时用脸部修复,
建议只开高清修复,
可用红色主题,黄色主题,绿色主题等控制画面色调,Yellow theme,red theme, green theme,  
或者混搭 黄蓝主题  Yellow and blue themes,红绿,red and green themes, 还有其它的自测哈

~~~~~~~~~~~~~~~~~~~~~~~o(* ̄︶ ̄*)o~~~~~~~~~~~~~~~~~~~~~~~

A Niji-style warrior model is available for use, and it has wide applicability, as it can be combined with various 2D base models.
The posture of holding a sword is very complex, so the chance of producing good hand poses is extremely low. However, it can be used with ControlNet for better results. In the future, a version without a weapon may be released if time permits.
The images are trained with simple backgrounds. If you want to use complex backgrounds, you can try using low weights and then increasing the weight of the background, or use layered LORA.
The following are the base models I have tested: DalcefoPainting_3rd, Duchaitenniji_cuteV235, and others that I have tested myself.
DalcefoPainting_3rd link: https://ko-fi.com/s/9dd78df21d
The embeddings models are: EasyNegative, ng_deepnegative_v1_75t, and badhandv4.
The VAE models are: vae-ft-mse-840000 and animevae. If the color of the base model is too heavy, you can switch to a VAE model or not use it at all.
I recommend using weights of 0.6-0.8 and using layered LORA to reduce complex effects. It depends on personal preference.
The sampler is DPM++ SDE Karras, and the CFG scale is 7.
If you want to produce a full-body image, you need to use a large-sized image; otherwise, the face will break. I am using 512x768, and after high-definition repair, it can be enlarged by 2 times to 1024x1536 or larger. Do not use face repair at the same time. I suggest using only high-definition repair.
You can use different color themes to control the color tone of the image, such as yellow theme, red theme, green theme, or mix and match different color themes, such as yellow and blue themes, or red and green themes. You can also try others according to your personal preference.

~~~~~~~~~~~~~~~~~~~~~~~o(* ̄︶ ̄*)o~~~~~~~~~~~~~~~~~~~~~~~

示例提示词和参数:
Prompt: ,weapon, 1girl, long hair,male focus, sword, solo, japanese clothes, holding, holding weapon, katana, holding sword, sheath, , , long sleeves, , haori, , hakama, kimono, sheathed,full body,
,, <lora:niji_wushi_yuan:0.6>

Negative prompt: easynegative

Steps: 20, CFG scale: 7, Sampler: DPM++ SDE Karras, Seed: 399951284, ENSD: 31337, Size: 512x768, Model: duchaitenniji_cuteV235, Model hash: 580174dabb, Hires upscale: 2, Hires upscaler: Latent, Denoising strength: 0.54, Clip skip: 2

本帖子中包含更多资源

您需要 登录 才可以下载或查看,没有帐号?立即注册 微信登录 手机动态码快速登录

x
回复

使用道具

 
您需要登录后才可以回帖 登录 | 立即注册 微信登录 手机动态码快速登录

本版积分规则

快速回复 返回顶部 返回列表