ForgeUI中使用SD3.5大模型

原创已于 2025-05-04 16:21:31 修改

· 3.4k 阅读

9 ·

版权

文章标签：

#stable diffusion #AI作画 #python #人工智能作画 #midjourney #AIGC

于 2024-11-03 18:20:45 首次发布

AIGC 专栏收录该内容

3 篇文章

订阅专栏

1 引言

Stability AI 发布了SD3.0的后续版本，即SD3.5大模型。ComfyUI已支持该模型，ForgeUI还需等待后续更新。本人在某管（传送门：SD3.5 EN FORGE WEBUI，博主是西班牙的Academia SD，有上网条件的朋友可以直接观看他的视频，如果能订阅点赞最好啦）找到一种在当前ForgeUI抢鲜体验SD3.5的方法，在这里也分享给大家。

CK	参数	大小	备注
SD3.0	20亿	\	只能说差强人意
SD3.5 Large	80亿	15G	商用友好听说不错准备试试估计效果会在Flux schnel和FP8之间
SD3.5Large Turbo	80亿	15G	商用友好听说不错准备试试估计效果会在Flux schnel和FP8之间
Flux.1(FP8)	120亿	11G	dev商用不友好 Flux schnel效果打折扣

B站发布了大量Flux与SD3.5的对比测评，这里推荐几个（如果可以，请给各位UP一键三连哦）

1.SD3.5深度评测全方位对比FLUX模型

2. 重磅更新！SD3.5！真正的开源可商用dit模型

3.SD3深度评测对比Midjourney

2 在ForgeUI中使用SD3.5

话不多说，直接上干货。

2.1 准备工作

（0）更新ForgeUI

使用原生UI的读者，请运行update.bat文件（不推荐，更新后也可能不能成功使用SD3.5）
使用秋叶启动的读者（非常推荐），请更新到2024.10.02发布的开发版本（其他版本可能不能成功使用SD3.5）

（1）下载SD3.5模型

在Liblib（推荐，速度很快）或者Hugging Face（需上网技巧）下载sd3.5_large_turbo.safetensors（4步快速出图版本，推荐）或者sd3.5_large.safetensors文件（ForgeUI使用这个模型效果不佳，没搞懂为什么，欢迎评论区交流）
将sd3.5模型放到ForgeUI目录的\models\Stable-diffusion文件夹下

（2）下载编码文件

可在Hugging Face（需上网技巧）下载clip_g.safetensors，clip_l.safetensors，t5xxl_fp8_e4m3fn.safetensors（也可下载t5xxl_fp16.safetensors，看自己配置自行决定就好）三个文件
将上述文件放置到ForgeUI目录的\models\text_encoder文件夹下

提示

一般来说，若之前使用过Flux大模型，这几个文件不需重复下载
clip_l.safetensors，t5xxl_fp8_e4m3fn.safetensors也可在Liblib下载，clip_g.safetensors（大小1.3G）真的没找到不需上网技巧和网盘的方法，请大家克服一下网盘限速或者直接在报脸下载，1.3G要不了太久。

（3）下载SD3.5的ForgeUI适配文件

下载由博主Academia SD提供的SD35_ForgeV2.zip文件（需上网技巧）
将上述压缩文件解压后得到backend、modules、repositories三个文件夹；然后将上述文件夹复制到forgeUI根目录（就是原本就含有backend、modules、repositories的位置，点击替换文件就ok）

（4）整合文件

若读者没有上网技巧或魔法值不足，也可下载我的百度盘整理包。网盘里面包含了sd3.5模型、编码文件，SD3.5适配文件，但编码文件没有t5xxl_fp16.safetensors，若确有需要，建议去报脸下载。

PS：这里就不提供ForgeUI的安装方式了哦，网上有很多教程，需要的小伙伴可自行寻找。

2.2 SD3.5使用教程

2.2.1 设置

完成上一节所有准备后，请参照以下设置：

UI选择all，checkpoint选择SD_3.5_L_Turbo.safetensors，VAE / Text Encoder选择clip_g.safetensors,clip_l.safetensors,t5xxl_fp8_e4m3fn.safetensors，采样器选择Euler/DPM++2M，调度器选择SGM Uniform（个人感觉采样/调度器选择这个搭配比较好），也可以选择Euler+beta
CFG scale设置为1（重要），不需要负面提示词
Distilled CFG Scale可设置为4（按需设置）
图片尺寸：按需设置
迭代步数：只需4步

2.2.2 一个例子

如果希望复刻我的图片，请按照以下设置：

正面提示词

A cute and charming girl wears a soft and fluffy Cat's ears (Steamed cat-ear shaped bread) headdress on her head. The inside of the cat ear is light pink,and the tip of the ear is white. Her face is delicate and small,her skin is fair and rosy,her eyes are deep and clear like lake water,and her eyelashes are long and curly. Her black and shiny hair draped over her shoulders,with a few strands resting on her cheeks. She was wearing a loose white sweater with a slightly open neckline revealing a slender neck. The sweater had snowflake like patterns and was paired with a light blue short skirt with lace trim. Her lips were like rose petals,slightly pouting with a smile,and she had silver earrings hanging from her ears.,
A blurred light gray space with subtle light and shadow transitions,resembling a simple indoor scene,but without many specific furniture or decorative elements,mainly to highlight the main characters.,
The overall color tone is mainly soft and elegant. The white sweater and light blue skirt of the main character laid the foundation of the light color system,and the light pink and white of the Cat's ears (Steamed cat-ear shaped bread) headdress were also soft colors. The light gray color tone of the background echoes the subject and does not grab their attention,creating a fresh,cute,and slightly dreamy color atmosphere throughout the entire picture.,

负面提示词：不填
UI选择all，checkpoint选择SD_3.5_L_Turbo.safetensors
迭代步数：4
VAE / Text Encoder选择clip_g,clip_l,t5xxl_fp8_e4m3fn
采样器选择Euler，调度器选择SGM Uniform
CFG scale设置1，Distilled CFG Scale设置4
图片尺寸：1024*1536
随机种：118341057987932
其他设置：默认
图片效果：

A cute and charming girl wears a soft and fluffy Cat's ears (Steamed cat-ear shaped bread) headdress on her head. The inside of the cat ear is light pink,and the tip of the ear is white. Her face is delicate and small,her skin is fair and rosy,her eyes are deep and clear like lake water,and her eyelashes are long and curly. Her black and shiny hair draped over her shoulders,with a few strands resting on her cheeks. She was wearing a loose white sweater with a slightly open neckline revealing a slender neck. The sweater had snowflake like patterns and was paired with a light blue short skirt with lace trim. Her lips were like rose petals,slightly pouting with a smile,and she had silver earrings hanging from her ears.,
A blurred light gray space with subtle light and shadow transitions,resembling a simple indoor scene,but without many specific furniture or decorative elements,mainly to highlight the main characters.,
The overall color tone is mainly soft and elegant. The white sweater and light blue skirt of the main character laid the foundation of the light color system,and the light pink and white of the Cat's ears (Steamed cat-ear shaped bread) headdress were also soft colors. The light gray color tone of the background echoes the subject and does not grab their attention,creating a fresh,cute,and slightly dreamy color atmosphere throughout the entire picture.,
Steps: 4, Sampler: Euler, Schedule type: SGM Uniform, CFG scale: 1, Seed: 118341057987932, Size: 1024x1536, Model hash: 6b6a0d8eb2, Model: SD_3.5_L_Turbo, Conditional mask weight: 1, Version: f2.0.1v1.10.1-previous-546-gf4d5e8ca, Module 1: clip_g, Module 2: clip_l, Module 3: t5xxl_fp8_e4m3fn

OK啦 hope you all enjoy SD3.5 in ForgeUI.