为什么有这么多图片压缩工具，还要再做一个？

[littletong]

最近频繁使用图片压缩功能，发现目前主流图片压缩工具还是有些不足：

• TinyPNG 会上传图片，且限制了 5M 大小。压缩后无法对比效果。
• Squoosh 不支持批量压缩图片，效率太低。且无法压缩到指定 KB 以内。

所以叒整了一个新的图片压缩工具，算是把几个痛点都解决了：

1. 隐私安全，不上传图片，压缩速度也成倍提升；
2. 批量压缩，不限制个数；
3. 大图片压缩，不限制大小；
4. 支持指定 KB 压缩；
5. 支持指定尺寸和压缩质量，图片格式；
6. 支持压缩后对比效果；
7. 默认高效的 WebP 格式，透明度+小+质量高，主流浏览器和系统都支持；

欢迎大家体验并提出改进意见 https://compressimage.cn

[iMusic]

第 4 点怎么做到的？

[zqjilove]

界面舒服，点个赞

[zqjilove]

给个建议，添加到列表里待处理的，无法删除，这个体验不够好。

[miaomiao888]

啊？
https://ibb.co/6vyjrQk

[li1218040201]

这类工具核心是压缩算法

刚刚试了下，一张 280kb 的 PNG 图片，你的压缩率 0 。。。TinyPNG 能压缩到 76KB

https://image.988589.xyz/api/cfile/AgACAgUAAyEGAASC5r41AAMoZ1AdKZVcixStxWYgMs8oBpLgC8IAAtq9MRsNv4FW6VqTga4eZuIBAAMCAAN5AAM2BA

[li1218040201]

@miaomiao888 笑死

[bugoftime]

👍。试了下，保持原格式 png ，80p 质量，变大了好多啊

[ypzhou]

@li1218040201 有原图吗 我想试下我的压缩工具

[li1218040201]

@ypzhou https://image.988589.xyz/api/cfile/BQACAgUAAyEGAASC5r41AAMqZ1Aeb8Tw5bKfYfz_Suzm8Eryae4AAtYUAAINv4FW2lZErBoxsvw2BA

[Outshine]

页面很清爽干净啊。

你这 cn 域名放 vercel ？

不考虑盈利？

[siweipancc]

移动端不友好啊

[lizhenda]

cn 域名不太好做海外吧

[liu731]

试了一下，建议别叫压缩工具，叫图片格式转换算了

[chutsetien]

首先，webp 的画质非常差，用 cwebp -mt -q 100 -m 6 -sharp_yuv -af -alpha_filter best -sns 0 压出来的 webp 都会与原图有放大后能够看到的明显区别，毕竟是 VP8 打底，最高水准也就那样了。当然，确实，体积小。
目前建议 jpegli. jpegli 的 q 90 感觉和上面那个各项打满的 q 100 的质量差不多。当然要求高一些的话可以往 q 95+ 走。
我有个自用的 batch 档，也是做转换和切割图片的——我不是 programmer, 真不是，而且 batch 纯自用，因此各种不足之处还请见谅了。不知道 V 站怎么贴，就试试看 markdown 了。
```
@echo off
chcp 65001
setlocal enabledelayedexpansion
set "a="
set "i="
set "fn="
set "bn="
set "on="
set "gn="
set "np="
set "op="
set "cm="
set "sm="
set "hg="
set "jq="
set "wq="
set "sl="
set "sp="
set "ss="
set "sx="
set "sy="
set "cr1="
set "cr2="
set "cx1="
set "cx2="
set "sf1="
set "sf2="
set "cmb="
set "smb="
set "vfb="
set "cpf="
set "spf="
set "jpg="
set "png="
set "wbp="
set "sns="
:a
if "%~1"=="" goto p
set "a=%~1"
if "!a:~0,1!"=="-" (
if "!a!"=="-op" (set "np=%~2" & shift)
if "!a!"=="-on" (set "gn=%~2" & set "hg=1" & shift)
if "!a!"=="-ob" (set "bn=%~2" & shift)
if "!a!"=="-p" (set "png=1")
if "!a!"=="-wpq" (set "wbp=1" & set "sns=80" & set "wq=80")
if "!a!"=="-wlq" (set "wbp=1" & set "sns=50" & set "wq=85")
if "!a!"=="-wmq" (set "wbp=1" & set "sns=20" & set "wq=90")
if "!a!"=="-w" (set "wbp=1" & set "sns=0" & set "wq=95")
if "!a!"=="-whq" (set "wbp=1" & set "sns=0" & set "wq=99")
if "!a!"=="-jpq" (set "jpg=1" & set "jq=80")
if "!a!"=="-jlq" (set "jpg=1" & set "jq=85")
if "!a!"=="-jmq" (set "jpg=1" & set "jq=90")
if "!a!"=="-j" (set "jpg=1" & set "jq=95")
if "!a!"=="-jhq" (set "jpg=1" & set "jq=99")
if "!a!"=="-rg" (set "cm=rg" & set "cpf=rgr")
if "!a!"=="-cx" (set "cm=cx" & set "cx1=%~2" & set "cx2=%~3" & set "cpf=wth!cx1!t!cx2!" & shift & shift)
if "!a!"=="-cr" (set "cm=cr" & set "cr1=%~2" & set "cr2=%~3" & set "cpf=lts!cr1!t!cr2!" & shift & shift)
if "!a!"=="-gy" (set "cm=gy" & set "cpf=grv")
if "!a!"=="-gx" (set "cm=gx" & set "cpf=grh")
if "!a!"=="-gr" (set "cm=gr" & set "cpf=gr")
if "!a!"=="-sq" (set "cm=sq" & set "cpf=sq")
if "!a!"=="-up" (set "sm=up" & set "sp=%~2" & set "spf=!sp!pc" & shift)
if "!a!"=="-uf" (set "sm=uf" & set "sf1=%~2" & set "sf2=%~3" & set "spf=!sf1!d!sf2!" & shift & shift)
if "!a!"=="-uy" (set "sm=uy" & set "sy=%~2" & set "spf=x!sy!" & shift)
if "!a!"=="-ux" (set "sm=ux" & set "sx=%~2" & set "spf=!sx!x" & shift)
if "!a!"=="-us" (set "sm=us" & set "ss=%~2" & set "spf=s!ss!" & shift)
if "!a!"=="-ul" (set "sm=ul" & set "sl=%~2" & set "spf=l!sl!" & shift)
if "!a!"=="-sp" (set "sm=sp" & set "sp=%~2" & set "spf=!sp!pc" & shift)
if "!a!"=="-sf" (set "sm=sf" & set "sf1=%~2" & set "sf2=%~3" & set "spf=!sf1!d!sf2!" & shift & shift)
if "!a!"=="-sy" (set "sm=sy" & set "sy=%~2" & set "spf=x!sy!" & shift)
if "!a!"=="-sx" (set "sm=sx" & set "sx=%~2" & set "spf=!sx!x" & shift)
if "!a!"=="-ss" (set "sm=ss" & set "ss=%~2" & set "spf=s!ss!" & shift)
if "!a!"=="-sl" (set "sm=sl" & set "sl=%~2" & set "spf=l!sl!" & shift)
) else (
if not defined i set "i=%~1"
for %%f in ("!i!") do set "op=%%~dpf" & set "on=%%~nf"
)
shift
goto a
:p
if "!np!"=="" set "np=!op!"
if not "!np:~-1!"=="\" set "np=!np!\"
if not exist "!np!" md "!np!"
if "!bn!"=="" set "bn=!on!"
if defined cpf set "cpf=-!cpf!"
if defined spf set "spf=-!spf!"
if "!gn!"=="" set "gn=!bn!!cpf!!spf!"
if defined cm (
if "!cm!"=="cx" set "cmb=crop='floor(if(gte(iw/ih,!cx1!/!cx2!),ih*!cx1!/!cx2!,iw)/2)*2:floor(if(gte(iw/ih,!cx1!/!cx2!),ih,iw*!cx2!/!cx1!)/2)*2'"
if "!cm!"=="cr" set "cmb=crop='floor(if(gte(iw,ih),min(iw,ih*(!cr1!/!cr2!)),min(iw,ih*(!cr2!/!cr1!)))/2)*2:floor(if(gte(iw,ih),min(ih,iw*(!cr2!/!cr1!)),min(ih,iw*(!cr1!/!cr2!)))/2)*2'"
if "!cm!"=="rg" set "cmb=crop='floor(if(gte(iw,ih),min(iw,ih*(17711/28657)),min(iw,ih*(28657/17711)))/2)*2:floor(if(gte(iw,ih),min(ih,iw*(28657/17711)),min(ih,iw*(17711/28657)))/2)*2'"
if "!cm!"=="gy" set "cmb=crop='floor(if(gte(iw/ih,17711/28657),ih*17711/28657,iw)/2)*2:floor(if(gte(iw/ih,17711/28657),ih,iw*28657/17711)/2)*2'"
if "!cm!"=="gx" set "cmb=crop='floor(if(gte(iw/ih,28657/17711),ih*28657/17711,iw)/2)*2:floor(if(gte(iw/ih,28657/17711),ih,iw*17711/28657)/2)*2'"
if "!cm!"=="gr" set "cmb=crop='floor(if(gte(iw,ih),min(iw,ih*(28657/17711)),min(iw,ih*(17711/28657)))/2)*2:floor(if(gte(iw,ih),min(ih,iw*(17711/28657)),min(ih,iw*(28657/17711)))/2)*2'"
if "!cm!"=="sq" set "cmb=crop='floor(min(iw,ih)/2)*2:floor(min(iw,ih)/2)*2'"
)
if defined sm (
if "!sm!"=="sp" set "smb=scale='if(lte(!sp!,100),floor(iw*!sp!/100),iw):if(lte(!sp!,100),floor(ih*!sp!/100),ih)'"
if "!sm!"=="sf" set "smb=scale='if(lte(iw*!sf1!/!sf2!,iw),floor(iw*!sf1!/!sf2!),iw):if(lte(ih*!sf1!/!sf2!,ih),floor(ih*!sf1!/!sf2!),ih)'"
if "!sm!"=="sy" set "smb=scale='-2:if(lte(!sy!,ih),!sy!,ih)'"
if "!sm!"=="sx" set "smb=scale='if(lte(!sx!,iw),!sx!,iw):-2'"
if "!sm!"=="ss" set "smb=scale='if(lt(iw,ih),min(!ss!,iw),-2):if(gte(iw,ih),min(!ss!,ih),-2)'"
if "!sm!"=="sl" set "smb=scale='if(gte(iw,ih),min(!sl!,iw),-2):if(lt(iw,ih),min(!sl!,ih),-2)'"
if "!sm!"=="up" set "smb=scale='floor(iw*!sp!/100):floor(ih*!sp!/100)'"
if "!sm!"=="uf" set "smb=scale='floor(iw*!sf1!/!sf2!):floor(ih*!sf1!/!sf2!)'"
if "!sm!"=="uy" set "smb=scale=-2:!sy!"
if "!sm!"=="ux" set "smb=scale=!sx!:-2"
if "!sm!"=="us" set "smb=scale='if(lt(iw,ih),!ss!,-2):if(gte(iw,ih),!ss!,-2)'"
if "!sm!"=="ul" set "smb=scale='if(gte(iw,ih),!sl!,-2):if(lt(iw,ih),!sl!,-2)'"
)
if defined cm if not defined sm (set "vfb=!cmb!")
if not defined cm if defined sm (set "vfb=!smb!")
if defined cm if defined sm (set "vfb=!cmb!,!smb!")
if not defined vfb (
ffmpeg -i "!i!" -vframes:v 1 -update true -compression_level 0 "!op!!on!_intermediateagzk3kXe8nDr.png" -y
) else (
ffmpeg -i "!i!" -vframes:v 1 -update true -compression_level 0 -vf "!vfb!" -sws_flags lanczos+accurate_rnd+full_chroma_inp+full_chroma_int "!op!!on!_intermediateagzk3kXe8nDr.png" -y
)
if "!jpg!"=="1" (
cjpegli -q !jq! --chroma_subsampling=444 "!op!!on!_intermediateagzk3kXe8nDr.png" "!op!!on!_intermediateagzk3kXe8nDr.jpg"
exiftool -overwrite_original -all= "!op!!on!_intermediateagzk3kXe8nDr.jpg"
if "!hg!"=="1" (
move /y "!op!!on!_intermediateagzk3kXe8nDr.jpg" "!np!!gn!.jpg"
) else (
move /y "!op!!on!_intermediateagzk3kXe8nDr.jpg" "!np!!gn!-q!jq!.jpg"
)
)
if "!wbp!"=="1" (
cwebp -mt -q !wq! -m 6 -sharp_yuv -af -alpha_filter best -sns !sns! "!op!!on!_intermediateagzk3kXe8nDr.png" -o "!op!!on!_intermediateagzk3kXe8nDr.webp"
exiftool -overwrite_original -all= "!op!!on!_intermediateagzk3kXe8nDr.webp"
if "!hg!"=="1" (
move /y "!op!!on!_intermediateagzk3kXe8nDr.webp" "!np!!gn!.webp"
) else (
move /y "!op!!on!_intermediateagzk3kXe8nDr.webp" "!np!!gn!-q!wq!.webp"
)
)
if "!png!"=="1" (
optipng -o7 -zm1-9 "!op!!on!_intermediateagzk3kXe8nDr.png"
exiftool -overwrite_original -all= "!op!!on!_intermediateagzk3kXe8nDr.png"
move /y "!op!!on!_intermediateagzk3kXe8nDr.png" "!np!!gn!.png"
) else (
del /q "!op!!on!_intermediateagzk3kXe8nDr.png"
)
endlocal
```

[givebest]

Mac 上可以试下这个仅 3MB 的 App： https://hp60.com/image2webp/

[chutsetien]

@chutsetien 顺附自写的该 batch 档的用法解释：

A tiny little Windows Batch script used for the cropping, scaling, and format conversion of images. It requires the following software to be properly set up in %PATH% (here in my case is C:\_c\):

ffmpeg;
libwebp;
optipng; and
exiftool.

Understanding ‘orientation-adaptive’ and ‘orientation-irrelevant’

As the script is used for the cropping and scaling of images, it would require the clarification of two concepts: ‘orientation-adaptive’ and ‘orientation-irrelevant’.

The former, as its name indicates, takes the image’s orientation into account. This can be easily understood in the context of scaling, as we often do need to scale an image by its ‘long edge’ or ‘short edge’, regardless of whether it is the width or the height.

It might get a little bit confusing when it comes to cropping. In cropping, by saying ‘orientation-adaptive’, it means that when given a ratio, the long edge always takes the first number, while the short edge takes the second. For example, if we have a horizontal image (which means its width is greater than its height), and we want to crop it to 4:3 orientation-adaptively, then the long edge, the width, will take the ‘4’ while the short edge, the height, will take the ‘3’. And when done, it would indeed be an image of 4:3. But if we have a vertical image (which means the height is the longer edge), and we still want to orientation-adaptively crop it to 4:3, then the height of the image, now the long edge, shall take the ‘4’, while the width of the image will take the ‘3’, and the final result we get is actually an image of 3:4.

What makes the ‘orientation-adaptive cropping’ more confusing is that it can be used to crop ‘against’ the image’s orientation by setting a ratio smaller than 1, in that it can force the long edge to take the smaller number in the ratio to become the new short edge while the original short edge can take the larger number and becomes the new long edge.

The ‘orientation-irrelevant’ scaling or cropping is much easier to understand—as its name indicates, it does not take the image’s orientation into account, only focusing on scaling or cropping based on the width and height, regardless of which one is longer.

Cropping switches

The script accepts switches used for cropping in an order of precedence as follows—if more than one switch is detected in one operation, a switch with a higher rank surpasses the lower ones:

-sq: crops the image to a 1:1 square;
-gr: orientation-adaptively crops the image to the golden ratio;
-gx: orientation-irrelevantly crops the image to the golden ratio by setting its width as the long edge;
-gy: orientation-irrelevantly crops the image to the golden ratio by setting its height as the long edge;
-cr x y: orientation-adaptively crops the image to a set ratio of x:y, where the long edge of the image takes the x, and the short edge takes the y;
-cx x y: orientation-irrelevantly crops the image to a set ratio of x:y, where the width of the image takes the x, and the height takes the y; and
-rg: performs a ‘reversed’ orientation-adaptive golden ratio cropping, where the long edge of the image takes the smaller number of the golden ratio, and the short edge takes the larger (not very useful actually, and that’s why it gets the lowest precedence).

Scaling switches

The script accepts switches used for scaling in an order of precedence as follows—if more than one switch is detected in one operation, a switch with a higher rank surpasses the lower ones:

-sl x: proportionally scales the image by setting its long edge to x;
-ss x: proportionally scales the image by setting its short edge to x;
-sx x: proportionally scales the image by setting its width to x;
-sy x: proportionally scales the image by setting its height to x;
-sf x y: proportionally scales the image by a given fraction of x/y; and
-sp x: proportionally scales the image by a given percentage x (‘%’ is not needed when giving the percentage number).

The script, by default, prohibits upscaling, unless the user explicitly orders the script to do so by changing the first letter of these scaling switches to ‘u’ (-ul, -us, -ux, etc). I strongly dissuade anyone who intends to use this script to upscale any image, for that image upscaling simply shall not be done in this way as it needs much more complicated algorithms and should be done with dedicated manual tuning.

Output file format and quality switches

The script accepts switches -p, -j, and -w to indicate an output of a .png image, a .jpg image, or a .webp image. These switches can be used simultaneously to output more than one image. For -j and -w, there are four quality suffixes, hq (‘highest quality’), mq (‘medium quality’), lq (‘low quality’), and pq (‘poor quality’), each representing a fixed quality parameter (or a group of fixed quality-related parameters in the case of -w), that can be used by immediately appending to the letter j or q, with no spaces in between, to form a five-tier image quality levels (together with the two default ones of no sticking letters, the -j and -w, which are both ranked second best in the hierarchy) for the user to choose. And by knowing me, you can rest assured that even by using -jpq or -wpq, the output will still be acceptable. But since I use ffmpeg to encode JPEG files—which uses mjpeg as its encoder—the output files will then always be added with a ‘dash’ label in the names, indicating that they are intended for quick sharing uses, and are not expected to be used in serious scenarios. If -p is detected, then the output .png file will be compressed by optipng with maximum tries, which means it will take a considerably long period of time if the image is large. Lastly, all of the output files will be cleaned of EXIF information in the end.

Designating output filename and path

By default, the output file or files will be saved in the same location as the input file, using the original filename as the basename with text labels of the process or processes being done to it and of the image quality parameter appended in the end. And yet the output filename and path are designatable with the switches listed below:

-op "path\to\the\designated location\": if used, the output files will then be exported to the designated location;
-ob new_basename: if used, then the given string will surpass the file’s original name and be used as the designated basename—the text labels for processed and image quality will still be appended to the given basename; and
-on "new filename": if used, then the given string will be used as the full filename of the output file, with no more appendings.

Since one operation of the script can produce more than one file in different formats, strings given after -ob and -on shall not contain a file extension, as it will be added automatically according to the output format.

[chutsetien]

@chutsetien 虽然知道可能没什么人看，但是加注一下：

> But since I use ffmpeg to encode JPEG files—which uses mjpeg as its encoder—the output files will then always be added with a ‘dash’ label in the names, indicating that they are intended for quick sharing uses, and are not expected to be used in serious scenarios.

这句是老版本里的，现在的这个脚本用的是 cjpegli, 不是 ffmpeg 的 mjpeg 了。

[vsean]

compressimage + cn 这个域名搭配，不知道该说什么好

[li1218040201]

@chutsetien 贴个文件链接就好了，回复越长消耗铜币数越多的，而且确实太长了。。

[chutsetien]

@li1218040201 在我的个人 Wiki 上，因此不太方便贴。我真不是程序员，就是个从 DOS 一路走来的老一代电脑使用者，所以没有把这些放在 GitHub 之类的地方，自己的小站也是自娱自乐，就不献丑了 ：）

[icyalala]

我还是推荐一下 https://imageoptim.com/mac
本地批量压缩，也支持有损和无损，而且会用多种算法压缩最终选择体积最小的

[shyrock]

op 这不是个网络服务吗？怎么说不上传图片？

[LogicDX342]

@chutsetien #20 可以试试 https://pastebin.com/

[vampuke]

@shyrock #22 在浏览器就压缩了，不用上传到服务器

[anzu]

我在做漫画汉化，有时把 jpg 原图转换成 webp 后，80%质量的 webp 文件大小反而比无损 webp 大几倍。问了 AI 说可能是原图片有大量重复图案导致的。我猜测应该是网点的原因，所以要不要有损压缩还需要自己判断一下。

[chutsetien]

@LogicDX342 所以说我真不是程序员哈哈，之前都没听说过这个~谢谢推荐。

[BYDH]

市面上缺少的是能把每一张图片的有效视觉信息损失自动控制在及格线上的工具，需要做一些算法上的工作和大量的试验

[longsays]

https://i0.du0.org/i/4sd823.jpg
这张图片也是越压越大

[kalman03]

压缩一次有新的释义了

[BYDH]

@longsays 还真别说，我觉得这张图没法直接压，因为它相较于刚扫描出来的原图，已经压得很极限了，不改变分辨率、不涂抹纹理的情况下把这张图二压到更小的体积，细节的损失绝对是不可接受的。只能说用更先进的编码器去处理尚且无损的原图，或许会比这张更好

[Tiande]

Windows 下推荐 XnConvert ，好用，绝绝子

[hoythan]

从来不相信有慈善家，你这个项目看起来没法盈利，说明这个工具没法持久。

[mayli]

@BYDH 有的 jpg 80%就是，当然及格是个主管概念，有些人觉得 100%也不及格
有些人觉得 75%也行。
微信压缩图片挺狠的，也没挡住它的流行

[BYDH]

@mayli 问题就在这里了，有的是 q=80 ，那还有的呢？一张满是纹理的图，和一张背景虚化的特写，最合适的质量参数绝对是不一样的，因为纹理一般认为不可舍弃，放大很容易看出端倪，而虚化的背景就无所谓。我认为可以做个算法，去衡量人眼会关注的局部区域(视觉显著性/边缘检测之类)相对于原图的质量损失，让程序算出一个"主观数据"，反馈到输入端，用于调整 q 。给用户喂💩是厂商的行为，用户没有必要亏待自己，在算力充裕的情况下，拿时间换无脑但稳定的输出是可行的，至少应该是可选的，然而并没有这样的工具

[mayli]

@BYDH 我觉得是你在民科了，jpg 保留的主要就是人眼主要关注的区域，而且对于压缩后的损失算法也有大量现成的比如 ssni/vmaf 。
Jpg 的 q 就是质量参数了，你复杂图和简单色块在 q=80 的时候压缩出来的体积明显差很多。
建议民科不要没事想着做个算法，现有的工具已经足够成熟了，而且除了成熟还有效率和兼容性的权衡。
你要是觉得有损压缩是吃屎，选择无损图片就好了，png webp av1 都有，就是相对于 jpg 都是小众需求。