5分钟系列：5分钟学会2023年最火的AI绘画（性能倍增：Stable Diffusion torch2.0.0 + cu118）

2023-05-02_5分钟学会2023年最火的AI绘画（性能倍增：Stable Diffusion torch2.0.0 + cu118）

系列文章目录请看第一篇：

2023-03-10_5分钟学会2023世界顶级AI绘画神器Stable Diffusion（入门篇）

交流可加扣裙：四二642八九七2

在之前的版本中，一直使用的是Torch 1.x版本，并且A1111目前为止已经大概一个月没有更新webui了，在迭代的过程中出现了一堆bug导致笔者一直没用最新版，比较明显的bug比如Lora选择框不能选择None，所以如果有Lora选择框就必须选择一个Lora，这显然是不符合使用逻辑的，在github上也有一堆issue，无人解决，因为提交的一堆Pull Request无人合并，只有原作者有权限，所以有人fork了一个新的分支vladmandic的WebUI，使用了Torch 2.0

• https://github.com/vladmandic/automatic

但是就在昨天（20230501），原作者终于回归了，并且发布了一个新的版本，详细如下

• https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases

具体更新内容如下

Features:

• switch to torch 2.0.0 (except for AMD GPUs)
• visual improvements to custom code scripts
• add filename patterns: [clip_skip], [hasprompt<>], [batch_number], [generation_number]
• add support for saving init images in img2img, and record their hashes in infotext for reproducability
• automatically select current word when adjusting weight with ctrl+up/down
• add dropdowns for X/Y/Z plot
• setting: Stable Diffusion/Random number generator source: makes it possible to make images generated from a given manual seed consistent across different GPUs
• support Gradio’s theme API
• use TCMalloc on Linux by default; possible fix for memory leaks
• (optimization) option to remove negative conditioning at low sigma values #9177
• embed model merge metadata in .safetensors file
• extension settings backup/restore feature #9169
• add “resize by” and “resize to” tabs to img2img
• add option “keep original size” to textual inversion images preprocess
• image viewer scrolling via analog stick
• button to restore the progress from session lost / tab reload

Minor:

• gradio bumped to 3.28.1
• in extra tab, change extras “scale to” to sliders
• add labels to tool buttons to make it possible to hide them
• add tiled inference support for ScuNET
• add branch support for extension installation
• change linux installation script to insall into current directory rather than /home/username
• sort textual inversion embeddings by name (case insensitive)
• allow styles.csv to be symlinked or mounted in docker
• remove the “do not add watermark to images” option
• make selected tab configurable with UI config
• extra networks UI in now fixed height and scrollable
• add disable_tls_verify arg for use with self-signed certs

Extensions:

• Add reload callback
• add is_hr_pass field for processing

Bug Fixes:

• fix broken batch image processing on ‘Extras/Batch Process’ tab
• add “None” option to extra networks dropdowns
• fix FileExistsError for CLIP Interrogator
• fix /sdapi/v1/txt2img endpoint not working on Linux #9319
• fix disappearing live previews and progressbar during slow tasks
• fix fullscreen image view not working properly in some cases
• prevent alwayson_scripts args param resizing script_arg list when they are inserted in it
• fix prompt schedule for second order samplers
• fix image mask/composite for weird resolutions #9628
• use correct images for previews when using AND (see #9491)
• one broken image in img2img batch won’t stop all processing
• fix image orientation bug in train/preprocess
• fix Ngrok recreating tunnels every reload
• fix –realesrgan-models-path and –ldsr-models-path not working
• fix –skip-install not working
• outpainting Mk2 & Poorman should use the SAMPLE file format to save images, not GRID file format
• do not fail all Loras if some have failed to load when making a picture

感兴趣可以自行读一下更新内容，其中最重要的莫过于使用了torch 2.0.0版本，性能方面有了比较大的提升，笔者测试下来，甚至提升了将近一倍，出图时间缩小为原来的1/2，下面详细测试一下

安装

建议直接克隆官方仓库即可，但是注意记得原来文件的备份，建议将所有模型等大文件单独放置一个目录，使用软连接的方式将多个WebUI链接到同一个模型文件目录上，节省硬盘空间，在Windows中，使用如下方式进行软连接

mklink /D  F:AIstable-diffusion-webui-betamodels   D:AIlarge_modelsmodels
mklink /D  F:AIstable-diffusion-webui-stablemodels   D:AIlarge_modelsmodels

1、将D:AIlarge_modelsmodels目录软连接至目录F:AIstable-diffusion-webui-betamodels

2、将D:AIlarge_modelsmodels目录软连接至目录F:AIstable-diffusion-webui-stablemodels

这样beta和stable都可以使用同一个模型目录！

测试图片

• Promot

best quality,(masterpiece:1.2),photo realistic, raw photo, (fractal art:1.2),colorful,(abstract water color background:1.3), soothing tones,blue eyes,blue hair,
1 female, mature female, (koh_amberheard:0.8), looking at viewer, hanfu, ming style, <lora:hanfu_v30:0.4>
Negative prompt: easynegative,badhandv4,watermark,text,black and white photos,(worst quality:1.5), (low quality:1.5), (normal quality:1.5), lowres, bad anatomy, bad hands, normal quality, ((monochrome)), ((grayscale)),
best quality,(masterpiece:1.2),photo realistic, raw photo, (fractal art:1.2),colorful,(abstract water color background:1.3), soothing tones,
1 female, mature female, (koh_amberheard:0.8), looking at viewer, __my\hanfu4__, (folk dance pose),"
Steps: 12, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 902893637, Size: 512x768, Model hash: 4d91c4c217, Model: lyriel_v15, Denoising strength: 0.5, ENSD: 31337, Hires upscale: 2, Hires steps: 10, Hires upscaler: R-ESRGAN 4x+ Anime6B

• 模型：lyriel_v15
• VAE：vae-ft-mse-840000-ema-pruned

图片如下

测试方法

将以上图片在旧版本和新版本分别运行三次，观察内存和时间用量，取最大最小和平均值，基本测试后在测试放大4K，使用，使用ControNet的tile模型加上Tiled DIffusion放大的到4K，这个过程等待比较久，就测试一次了，放大4K参数如下，所有参数都在下面文本中，这里就不再截图赘述

(half portrait:1.3),
best quality,(masterpiece:1.2),photo realistic, raw photo, (fractal art:1.2),colorful,(abstract water color background:1.3), soothing tones,blue eyes,blue hair,
1 female, mature female, (koh_amberheard:0.8), looking at viewer, hanfu, ming style, <lora:hanfu_v30:0.4>
Negative prompt: easynegative,badhandv4,watermark,text,black and white photos,(worst quality:1.5), (low quality:1.5), (normal quality:1.5), lowres, bad anatomy, bad hands, normal quality, ((monochrome)), ((grayscale)),
best quality,(masterpiece:1.2),photo realistic, raw photo, (fractal art:1.2),colorful,(abstract water color background:1.3), soothing tones,
1 female, mature female, (koh_amberheard:0.8), looking at viewer, __my\hanfu4__, (folk dance pose),"
Steps: 12, Sampler: Euler a, CFG scale: 7, Seed: 902893637, Size: 2048x3072, Model hash: 4d91c4c217, Model: lyriel_v15, Denoising strength: 0.75, ENSD: 31337, Mask blur: 4, Tiled Diffusion upscaler: R-ESRGAN 4x+ Anime6B, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 128, 'Latent tile height': 128, 'Overlap': 16, 'Tile batch size': 8, 'Upscaler': 'R-ESRGAN 4x+ Anime6B', 'Scale factor': 2, 'Keep input size': True}", ControlNet-0 Enabled: True, ControlNet-0 Module: none, ControlNet-0 Model: control_v11f1e_sd15_tile [a371b31b], ControlNet-0 Weight: 1, ControlNet-0 Guidance Start: 0, ControlNet-0 Guidance End: 1
Used embeddings: easynegative [119b], badhandv4 [dba1], koh_amberheard [23d3]
Used embeddings: koh_amberheard [23d3]

后面发现这里少了个重要参数，Encoder Tile Size和Decoder Tile Size，这两个值分别是1536和96，根据事情情况设定，太大可能会爆显存。

这里贴一下硬件参数：

• 处理器 AMD Ryzen 5 2600X Six-Core Processor 3.60 GHz
• 机带 RAM 32.0 GB
• 显卡：4070TI

旧版本：torch1.13.1+cu117+xformers0.0.16

• 第一次测试结果

Time taken: 33.91sTorch active/reserved: 4929/6432 MiB, Sys VRAM: 8722/12282 MiB (71.01%)

• 第二次测试结果i

Time taken: 27.76sTorch active/reserved: 4925/6428 MiB, Sys VRAM: 8726/12282 MiB (71.05%)

• 第三次测试结果

Time taken: 27.82sTorch active/reserved: 4925/6428 MiB, Sys VRAM: 8734/12282 MiB (71.11%)

• 4K高清修复

Time taken: 1m 58.12sTorch active/reserved: 3303/3732 MiB, Sys VRAM: 11519/12282 MiB (93.79%)

新版本：torch2.0.0.+cu118+xformers0.0.18

• 第一次测试结果

Time taken: 21.47sTorch active/reserved: 5373/8704 MiB, Sys VRAM: 12257/12282 MiB (99.8%)

• 第二次测试结果i

Time taken: 15.98sTorch active/reserved: 5368/8704 MiB, Sys VRAM: 12127/12282 MiB (98.74%)

• 第三次测试结果

Time taken: 15.81sTorch active/reserved: 5368/8704 MiB, Sys VRAM: 11990/12282 MiB (97.62%)

• 4K高清修复

Time taken: 1m 17.42sTorch active/reserved: 6362/9418 MiB, Sys VRAM: 12282/12282 MiB (100.0%)

结论

文生图+高清修复

• 时间

结论：第一次由于需要加载模型等初始化操作，所以用时比第二、三次久一些，第二次和第三次基本差不多，所以就以第三次作为参考，整体新版本比旧版本速度快43%左右

• 显存

结论：可以看到显存使用上其实新版本使用得更多，不清楚新版本是不是以空间换时间

的做法，这里没有大量数据验证，只能得出一个简单结论，新版可能是显存能用多少用多少，具体需要更多验证。

4K高清修复

• 时间

结论：新版比旧版快约34%，（后面又测试了一次，第二次运行应该还要快一些）

• 显存

结论：可参考文生图+高清修复的简单结论