Checkpoint 存档

yuysjx

已于 2024-07-06 09:43:00 修改

阅读量1.8k

点赞数 41

文章标签：大数据 AI作画

于 2024-07-06 09:42:29 首次发布

本文链接：https://blog.csdn.net/yuysjx/article/details/140223499

版权

Top view, perspective composition

推荐搭配LoRA: Nijiexpress V1 V2，Xiaorenshu小人书，墨心moxin，沁彩

推荐Prompts：Fisheyes, masterpiece, top quality, best quality, official art, beautiful and aesthetic, animation

Neg_Prompts: EasyNegativeV2, ng_deepnegative_v1_75t, (low_quality:1.4), (worst_quality:1.4), (badhandv4:1.1),collage, artist_name, signature, artist_logo, watermark

（推荐提示词：

Fisheyes, masterpiece, top quality, best quality,

official art, beautiful and aesthetic, animation,

Traditional Chinese beauty,ink wash painting style,Beijing opera costume等...

欢迎自行添加元素或使用珍藏提示词~

负面提示词推荐：

Realistic,EasyNegative, (worst quality, low quality:1.4), (depth of field, blur: 1.2), (greyscale, monochrome:1.1), 3D face, cropped, lowres, text, (nsfw:1.3), (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((grayscale)), skin spots, acnes, skin blemishes, age spot, (ugly:1.331), (duplicate:1.331), (morbid:1.21), (mutilated:1.21), (tranny:1.331), mutated hands, (poor draw hands:1.5), blur, (bad anatomy:1.21), (bad proportions:1.331), bs, extra disfigured (1.331), (missing arms:1.331), (extra legs:1.331), (fused fingers:1.61051), (too many fingers:1.61051), (unclear eyes:1.331), lowers, bad hands, missing fingers, extra digit,bad hands, missing fingers, (((extra arms and legs)))), freckles,(white border:2),nsfw

或EasyNegativeV2, ng_deepnegative_v1_75t, (low_quality:1.4), (worst_quality:1.4), (badhandv4:1.1),collage, artist_name, signature, artist_logo, watermark）

（AWPortrait第一款基于SDXL架构的大模型，覆盖不同人种，强大的语义理解能力，具有多场景、高质量、稳定的写实风格直出能力。增加了旗袍、汉服等东方元素的识别响应，相较于其他SDXL写实模型，在东亚人种的美观度上更高，更多模型效果请参考例图Prompts。

为了达到最好的生图品质，请遵照以下参数配置来进行设置。

采样推荐：DPM++ 2M Karras、Restart、Euler a ， Step 30

分辨率推荐：768*1152

放大算法：Lanczos，重绘幅度0.2-0.3

负面提示词：(worst quality),(low quality),(normal quality),lowres,bad anatomy,((bad hands,broken hands)),(many fingers),((grayscale)) watermark,moles,

Adetailer：

Close-up 特写情况下，可不开启或开启0.25-0.3的重绘幅度。

UpperBody 半身像情况下，可不开启或0.3-0.35的重绘幅度。

Fullbody 全身像情况下，请开启0.35-0.4的重绘幅度。）作者

（这是LarmeKei拉美风格大模型，融合CounterfeitV30_v30大模型，整体效果不错，欢迎尝鲜。

VAE选择：vae-ft-mse-840000-ema-pruned.ckpt

推荐反向提示词：verybadimagenegative_v1.3BadDreamFastNegativeV2badhandv4

推荐分辨率：768*512/512*768/768*768（建议开启高清修复，提升分辨率：放大两倍，放大算法：R-ESRGAN4x+Anime6B，重绘幅度：0.5,高分迭代步数开6~8）

采样方式：DPM++2MSDEKarras

迭代步数：35

提示词引导系数 (CFG Scale)：7~9）作者

看返图（好看）

LiblibAI·哩布哩布AI - 中国领先的AI创作平台

（使用介绍：

放大算法推荐 4x-UltraSharp 8x-NMKD-Superscale

迭代步数推荐 30左右

采样推荐 DPM++ 2M Karras DPM++ SDE Karras

VAE 840000-ema-pruned.safetensors

（使用介绍：

已经内置VAE

放大算法推荐 4x-UltraSharp 8x_NMKD-Superscale_150000_G

迭代步数推荐 30左右

原始尺寸推荐：768×1024 x1.5左右（太大容易错位）1024x1536

建议打开AD修脸功能: ADetailer face_yolov8n.pt

采样推荐 DPM++ 2M Karras

反向推荐 ng_deepnegative_v1_75t,badhandv4 (worst quality:2),(low quality:2),(normal quality:2),lowres,bad anatomy,bad hands,normal quality,((monochrome)),((grayscale)),wrong body structure,wrong limb,long body,deformed,malformed hands,poorly drawn hands,mutated hands,extra arms,extra limb,disconnected limbs,floating limbs,malformed limbs,too many fingers,extra legs,fused fingers,disfigured,Multiple people,bad body,）作者

(另外本模型所展示的风景类示意图，其中冰原极光场景的提示词来自博主Sam Hui，确实相当漂亮。以此为借鉴，我还生成了其他一些风光场景的示意图，在此感谢！

Updated Film Grian version 2.0. During the production process of this version, I conducted comparative tests by integrating Filmgirl Lora into the base model and using Filmgirl Lora's training set for Dreambooth training. Then I merged the two large models obtained, and carried out hierarchical weight adjustment. After multiple rounds of testing, Film Grain 2.0 performs better than 1.0 in terms of portrait aesthetics and skin tone.

Additionally, the landscape sketches displayed in this model, including the glacier aurora scene inspired by the keyword from blogger Sam Hui, are indeed quite beautiful. Taking this as a reference, I have also generated some other landscape sketches, and I would like to express my gratitude here!

更新了Film Grain 1.0版本。该版本采用film girl LoRA的训练集照片训练Dreambooth模型。本模型的真实感显著提升

Updated Film Grian version 1.0. This version uses the training set photos of film girl LoRA to train the Dreambooth model. The realism of this model has been significantly improved)作者

（基础tag：

•正：(masterpiece:1.2), (best quality:1.2),(high quality:1.2),(realistic:1.2), (photorealistic:1.2) ,(official art:1.2),(extremely detailed:1.2),1girl,1girl, （在此基础上添加）

wildcards,

•负：(worst quality:2),(low quality:2),(normal quality:2),lowres,watermark,(NSFW:1.2),(open mouth:1.6), (teeth:1.6),（可以不用负嵌入embeddings模型）

基础参数：

CLIP 终止层数：2

vae：已内置(可以叠加vae-ft-mse-840000）

采样：需要多样性随机性euler-a(20-30步)、DPM++ 2M Karras（20-30步）、restart（20-50步）、ddim（30-50步）、euler（20-50步）、DPM++3m系采样50步以上、DPM++2m系采样20-50步、其他采样亦可自行尝试

步数：20~50 （见上方采样推荐）

cfg：7

高分辨率修复参数：（一定要开！）（一定要开！）（一定要开！）重要的事情说三遍

高分迭代步数：0

放大算法：4x-UltraSharp(推荐使用）

重绘幅度：0.15~0.2（近景0.15，远景0.2）