GraphicsMagick Image Processing System | Hacker News Reader

GraphicsMagick Image Processing System

graphicsmagick.org

183 pointsbased246 comments

GraphicsMagick has been one of the hidden gems I've found over the years in my career. HN would probably mostly be interested in it for server-side applications [+], but it is an absurdly useful thing to have sitting around your laptop, too, for those times when e.g. you've got a screenshot of something but a government/etc site will only take a file upload in pdf format. "gm convert foo.png foo.pdf" and you're done.

[+] Really good for common "take an arbitrary user upload; grab info to crop it from your web app; crop and resize to appropriate dimensions" sorts of tasks.

The level of serendipity in HN is absurd. This is exactly what I was going to search for today. I'm going to use it local with Electron.

> I'm going to use it local with Electron

Interesting. I also have a use case for this though I wasn't sure how easy it would be to bundle something like GM with my Electron app. Have you any insights on this? I'm targeting both Win and Mac.

Windows can use the compiled binary, Mac will need a static build.

If you're using asar packages in your final app make sure it extracts those binaries then do a child_process.spawn

Thank you for the info. I'm looking for a link with all the details but I haven't found one so far. If you have that I would appreciate you sharing. Thanks!

If the app uses JS, I recently did a bunch of image processing using the node GM wrapper ('gm' on npm), and it was generally a huge pain in the ass. I found that I couldn't do the edits I needed in a single atomic call to gm, and I couldn't get the wrapper to pipe multiple calls together, so I wound up having to write and manage temp files, and it was all pretty hairy.

I later found a project called 'jimp', which is more or less similar to gm but written in pure JS, and which made it easy to chain calls together into nontrivial edits. Since it has no native dependencies it would probably be easier to bundle with electron as well. Presumably it's slower than gm of course (in my case that didn't matter, so I never looked back).

Now this is real serendipity! I was guessing that I'd have a difficult time with this integration. Jumping straight to your recommendation.

Well, YMMV but for me what took me several days with the gm wrapper, I had working in about 15 minutes with jimp. Now that I think about it, I recall that the gm wrapper had a very strange API, that greatly obscured what parameters got passed to gm underneath, and figuring out how that worked was a big part of the headache.

With jimp, everything basically did what it said on the tin. I remember wondering why such a useful library wasn't more famous.

Jimp is something I had come across as well and played with a bit. FYI I wrote in my notes "dependency on jpeg-js is dog slow." I recommend some perf tests for your use case before jumping too far in.

I (and a number of other folks) have also run into an enormous memory leak, apparently originating in jpeg-js, when using jimp. On the order of "loading a 3MB file leaks 200MB of memory", which makes it unusable if you're at all RAM-constrained. Shame, because it has a lot of promise otherwise.

Good to know. However I don't see any issues for this in the jpeg-js Github.[1]

The decoder.js script in jpeg-js is < 1000 loc, so I suspect this can be addressed if someone comes up with a reproducible case?

Edit: I'll happily try to do it if this isn't resolved by the time I get around to needing to use jimp / jpeg-js.

[1] https://github.com/eugeneware/jpeg-js/issues

On top of the memory issue, jimp format support seems limited. No gif support for example.

My personal favourite is node-canvas. It simply brings the HTML5 canvas api to node.

Does node-canvas allow you to select high fidelity algorithms over fast algorithms? A while ago, I wrote a crop tool using browser canvas, and found that the results of a scaling operation were unacceptable, due to its fast downsampling algorithm.

The gory details and a few attempts at higher quality scaling algorithm implementations are on SO: http://stackoverflow.com/questions/18922880/html5-canvas-res.... TL;DR: sending the image to the server and having GM scale it is faster and more reliable than your own scaling algorithm in single-threaded browser JS.

I think node canvas is a wrapper over cairo. It lets you set a quality value too. Whether cairo is comparable to GM is something I don't know.

Nice, I'm going to keep an eye on this, but I guess it will not be a huge problem because it will run single-user in a desktop machine.

You can drive graphicsmagick via stdin/stout if you start it with the batch command, and avoid writing temporary files.

Yes, in retrospect I think much of my difficulty came from the node 'gm' wrapper being hard to use - it would probably have been easier to invoke gm directly (as a child_process or however that's normally done).

And I was just debugging a "pixel cache allocation fail" bug with GM: https://github.com/aheckmann/gm/issues/598.

Personally I find that Apple Preview has always been good enough for these tasks. It can also do a surprising amount of PDF editing: for example, you can draw annotations on pages (for highlights and signatures), and selectively combine pages from multiple PDF documents into one by simple copy & paste.

"Personally I find that Apple Preview has always been good enough for these tasks."

OSS is a good choice because not every developer works on a Mac and with binary apps. You can hack quite a few quick/dirty tools to deal with repetitive tasks. Having said that Apple Preview is a pretty useful tool.

Except Preview 8 (post-Mavericks) sucks. Annotations of any kind always appear in the center of the page and have to be manually moved into position, it doesn't easily recognize selections, combined documents don't always retain changes, I could go on. I have a Mavericks VM just for Preview 7.

[edit] grammar

The issue with taking arbitrary user uploads and piping them directly into a program that happily accepts and interprets arbitrary data is security.

They claim support for over 200 formats, some of which surely have exciting scripting functionality that no one ever uses:

http://www.imagemagick.org/script/formats.php

And there have been recent code execution vulnerability reports:

https://imagetragick.com/

Now libpng (e.g.) has issues as well but if at all possible I would try to restrict the input format and process only with software for the specific format.

> HN would probably mostly be interested in it for server-side applications

This image processing system would be very useful to have on the client.

(Why can't we still use arbitrary stuff on the client that was written for the server?)

You can if you're shipping a native app. But if you're targeting the web you're limited by tools written in JS or what you can transcribe to JS.

asm.js and WebAssembly could very well change this in a few years.

To further clarify SIMD.js (along WebAssembly's SIMD.js hooks) will change this.

Right now you can still port GraphicsMagick to emscripten you'd just have to accept a performance hit on the parts that rely on SIMD.

Fyi... I last looked at both GraphicsMagick and ImageMagick in 2013 when I was writing a custom batch image converter. Because GraphicsMagick had a reputation for better performance[1] I initially wrote my wrapper tool around GraphicsMagick. However, I encountered bugs with Photoshop PSD files (e.g. not properly reading multi-layer PSD files.[2]) That's when I found out that ImageMagick tracks the PSD file format more reliably than GraphicsMagick. Yes, GM later fixed that particular bug in 2014 but I had already revised my tool to use IM.

Lesson learned: if processing PSD files is critical to your project, you'll want to favor ImageMagick over GraphicsMagick. The author of GM admits the parser of PSD is not the best so it's something to keep in mind.

[1]Flickr and Etsy used GM on their backends:

http://www.slideshare.net/jallspaw/operational-efficiency-ha...

http://www.web2expo.com/webexsf2009/public/schedule/detail/8...

[2]https://sourceforge.net/p/graphicsmagick/bugs/242/

[3]also some bonus links on why PSD is very difficult to parse correctly:

https://github.com/gco/xee/blob/master/XeePhotoshopLoader.m?...

http://blogs.adobe.com/jnack/2009/05/some_thoughts_about_the...

PSD support is now disabled by default in GM apparently: https://github.com/Homebrew/homebrew-core/issues/6072#issuec...

If you need to process huge images, GraphicsMagick does some things that ImageMagick can't do. A few years ago when I was making poster sized art prints, I would routinely process 60k^2 pixel (4 gigapixel) images. Just a resize with a Gaussian filter would thrash and never finish overnight in ImageMagick, and only took a minute or two in GraphicsMagick.

I'm not in love with the CLI for either, but GM's is better and I like that it's all collected under one roof rather than a whole bunch of separate commands.

I do wish the projects could merge back into one and have the best of both. I also wish they were better at color handling, resizing needs to be done at a higher bit depth, better support for converting between modern color spaces, etc.

I love sips on my mac because it's fast, but I'm not sure if anything's a better cross-platform choice for basic batch operations on images, or for the backend of a web service, than GraphicsMagick.

I'm curious, as I want to make some poster-sized art prints soon:

1) What resolution were the source files?

2) What upscaling options did you use?

3) What multiple of the original file size did you go up to? Were you limited by the quality of the upscaling, or did you just decide on the size of print and dpi, and multiply?

Oh, I was downscaling, mainly for filtering/anti-aliasing. My source files were the 60k^2 images, and the output was more like 20k^2... I shoot for at least 300 dpi and I was printing 60-inch prints.

I had enough trouble downscaling such large images that I ended up writing my own downscaling in my renderer so that I wouldn't have to use GraphicsMagick or anything else, but of everything I ever tried, GraphicsMagick was the best at handling enormous images.

FWIW, I've never been happy with any upscaling. If you do have the option to render at your target res and not use upscaling, that's ideal. If you must upscale, it's hard to find anything that can do more that 2-3x without noticeable artifacts or visible blurriness.

Thanks for the detailed reply. I'd assumed your source files were digital photos of artwork (e.g. 20 megapixel). I want to print some photos of Japanese wood block artwork, and there's an RNN implementation which apparently works well for doing 2x, 4x etc. upscaling: https://github.com/nagadomi/waifu2x

Ah, yes photos are difficult. I have used some of the pixel art upscalers, but not for prints, and I haven't used waifu2x before. They can be almost magical in some situations, and Japanese woodblocks seem like a pretty good application for waifu2x.

I'd worry a little about textured background and what happens in areas of high detail, but I'm speculating. You can see resolution artifacts in the waifu2x examples. But those source images are tiny -- it may behave completely differently on a 20MP image. It also depends on what material you print onto. Canvas and some popular art papers have a lot of texture that can hide the 1-2 pixel sized artifacts.

If you are taking the photos yourself, one option you might consider is to take overlapping photos of sub-portions of the woodblocks and stitch them together using a panorama editor like Hugin to get a source image that is at or above your target resolution.

If you have no other options but upscaling, do some experiments with the cheapest printing service you can find that will print 24x36 inches at 300dpi. There are a bunch online that will mail you prints very fast. If you're in the US, Costco does ridiculously cheap photo prints at 24x36. Experiment with upscalers and see what works best for the photos you have. You might find that you don't like any of them, or you might find out that you can't tell much difference between waifu2x and generic upscaling, or you might find out that waifu2x is perfect and unbeatable.

I was confused for a moment: Imagemagick and graphicsmagick are different, http://blog.trackduck.com/2014/11/12/imagemagick-vs-graphics...

GM is a fork. It's actually one of the more persistent and notable forks in the open source community. Both projects continue development and have no interest in reconciliation.

The last I heard about Image Magick was that it was out-dated and people were suggesting using other tools (though I don't remember what). Can anyone comment on that? This might have been only for a few specific tasks, not for Image Magick as a whole.

I don't know about now, but I've encountered significant issues with it every time I've attempted to use it in the last three years. To be fair, I was attempting to use it through wrappers - I had a particularly bad run trying to get around issues with both rmagick and minimagick in ruby, and later on a different project with wand in python. Though, with wand, the issues I encountered there were mostly with getting a buildpack working for it on Heroku.

Regardless, whether it's my incompetence or imagemagick itself, I don't have very good associations with the library.

GraphicsMagick fixed all the memory and CPU problems that ImageMagick was giving us in a large wiki farm production environment. That alone is a reason to switch to it.

do you have a covering blog post about that?

Unfortunately not. That occurred back under different management when talking about internal infrastructure changes was discouraged.

I switched to GraphicsMagick a few years ago too. I find that its command line options make more sense and that it produces better results. I do ad-hoc tests from time to time and and the difference between the two isn't earth shattering, but I get the feeling that GraphicsMagick is a better project for some reason (maybe it was something I read once).

Imagemagick does support some more advanced features such as deep colors, perceptually uniform color spaces (cielab) which graphicsmagick does not.

As I understand it some of the IM features are important for scientific image processing, HDRI or deep processing pipelines where rounding errors could accumulate.

GraphicsMagick is great, but it's not a replacement for ImageMagick. While it offers a few features not found in IM, it lags far behind on the more advanced stuff, such as image recognition/machine vision, and in general does not track the main project any longer.

vips is another great industrial-grade image processing library

https://github.com/jcupitt/libvips http://www.vips.ecs.soton.ac.uk/index.php?title=Libvips

And let's not forget NetPBM either, which (when I was doing this a decade ago) was faster and more memory efficient than (Image|Graphics)Magick.

CGAL https://news.ycombinator.com/item?id=12793220 (from a few days ago).

Options!

But GraphicsMagick and CGAL are nothing alike?

Reply on news.ycombinator.com