by David Tschumperlé

A new major version (numbered 3.0.0) of G'MIC (GREYC's Magic for Image Computing), an open source framework for image processing, was released on December 9th 2021. This project, distributed under the CeCILL free software license, is mainly developed by the IMAGE team of GREYC, a French research laboratory in Information and Communication Sciences and Technologies (Mixed Research Lab CNRS / ENSICAEN / University of Caen).


The last report on this software was published over two years ago, in August 2019. Since then many new features have been released. In this article, we detail some of the most significant ones.

It's time to make a hot drink, get comfortable on the couch and let us tell you about two years of development of a dynamic, free project!

Note: click on the images to get higher resolution versions. G'MIC is an open-source framework for digital image manipulation and processing.

It offers various user interfaces for the algorithmic manipulation (automated or semi-automated) of images and generic signals, ranging from simple 1D signals to 3D volumetric image sequences with any number of channels (which includes classical color images). The core of this project is based on the definition of a specialized scripting language (the "G'MIC language") and on a free implementation of its associated interpreter. This language was specifically developed to ease the prototyping and implementation of new image processing algorithms and operators. The users of the framework can thus apply operations among the hundreds already predefined, but also have the possibility if writing their own complex processing pipelines and make them accessible in the various user interfaces of the project. G'MIC is therefore, in essence, an open, extensible and constantly evolving framework.


The G'MIC project has been developed since 2008, mainly by two members of the IMAGE team of the GREYC laboratory: David Tschumperlé (CNRS research fellow, team leader) and Sébastien Fourey (ENSICAEN lecturer). The IMAGE team, one of the six teams of the GREYC, is composed of about fifty members (researchers, teacher-researchers, doctoral and post-doctoral students, engineers), all specialized in the fields of algorithmics and mathematics of image processing.

The most visible G'MIC user interfaces are: the gmic command, which can be used on the command line (an essential companion to ImageMagick or GraphicsMagick for those who process their images via the terminal), the G’MIC Online Web service, and above all, the G’MIC-Qt plugin, which is available for several popular digital image editing software (free or proprietary) such as GIMP, Krita, Paint.net, and more recently, Adobe Photoshop or Affinity Photo. This plugin allows you to enrich these programs with more than 570 different filters and effects to apply to images.


Thanks to its integrated scripting language, facilitating the prototyping and implementation of image processing algorithms and operators, new filters are regularly created and added to G'MIC. A large part of this article will be dedicated to the description of the new filters and effects developed in the last two years. G'MIC-Qt is the most popular and most used interface of the G'MIC project, so let's start by describing the new features concerning it. An exhaustive list of the new filters would be impossible here, since in fact more than fifty new entries have been added in two years! We have tried to follow the adage "A picture is worth a thousand words", by sprinkling this article with numerous figures illustrating the new features.



Here, everything is adjustable: the number of brush strokes at each iteration, the brush dynamics (size, opacity, color, orientation...), the geometric precision of the strokes, etc. This makes it a very versatile filter, allowing it to simulate various painting styles. This can be seen in the animation below, which shows the results obtained with several provided presets.








It is then easy, by mixing the result of this effect with the original image (through well chosen layer merging modes) to generate animations, which will transform any of your images into a horror show! (here we used it in conjunction with the G'MIC morph command to render the animation below).




Next, the Artistic / Stylize filter gets six new preset styles, developed by Christine Garner, an artist who had used the filter extensively to give her images a textured look before offering her contribution. These new styles allow, for example, the production of images with pencil, chalk, pastel or charcoal textures.




At first sight, this information may seem trivial, but it actually represents a milestone in the history of G'MIC (as well as many weeks of development!). Indeed, to implement this filter, we had to write a complete machine learning library from scratch, named nn_lib (for "Neural Network Library"). This library now allows the construction, learning, and evaluation of artificial neural networks (this type of method being generally referred to as "Artificial Intelligence"). The Repair / Denoise denoising filter is actually the first G'MIC filter based on this new machine learning library!


With the development of nn_lib, we hope to offer more and more neural network based filters for image editing and generation in G'MIC in the future. Great things to come!








For the color retouching of images, two new particularly interesting filters appear in G'MIC-Qt:

First, the Colors / Tune HSV Colors filter, which allows the user to define - in a very finely tuned way - a transformation function operating in the HSV (Hue-Saturation-Value) color space. This filter works in the following way:





From a technical point of view, this redefinition of the "key colors" leads initially to the construction of a dense interpolated CLUT (Color LUT) in 3D, which is then used to modify the input image according to the wishes of the user, as illustrated below in our example with the butterfly:


The use cases of this filter are quite large. Whether it is for a local color retouch (transformation of one color into another), or a more global colorimetric retouch, this filter will easily find its place in the toolbox of the color tweaker. In the figure below, we used this filter to define a twilight color atmosphere.


A demonstration video of this filter can be seen on the Youtube channel of Pierre "El Lobo", who regularly posts tutorials in French on how to use GIMP. Click on the image below to access it. More details on how to use this filter can also be found on the G'MIC forum.

The second notable colorimetric retouching filter is called Colors / Transfer Colors [PCA]. It makes it possible to transfer the colors of a reference image to a target image which one wishes to modify. Technically, this transfer is made by imposing the covariance matrix of the image to be modified to match that of the reference image.


Note that the underlying color transfer algorithm is not only used for this filter, but has also been integrated as a sub-part of more complex processes (e.g. the Artistic / Stylize filter mentioned earlier).

Lovers of image distortion and Glitch art may also be interested in this new version 3.0 of G'MIC.



Below, David Revoy demonstrates an application of this image break filter in the context of illustration, for a rendering of magical lightning with a most beautiful effect!


A more detailed video of David Revoy using this type of filter can be seen at the following link:



It is funny to notice that in the example above, the blocks corresponding to the eyes have simply been reversed in the two images!




Now let's enter the wonderful world of rendering filters. This is a unique category where you can find all sorts of oddities, filters that seem to have no use at all, until one day they become perfectly indispensable… or not! The following filters, which appeared in G'MIC 3.0, are no exception to this rule. They may be of little use to you, but they are still nice!







There are a lot of parameters, so this filter offers a set of presets to play with all possible curve variations. The figure below illustrates some of the predefined shapes offered by the algorithm.


As you can see, this filter allows you to draw a wide variety of parameterized sine curves. It can be used to generate shimmering backgrounds. Pierre "El Lobo" has also made a short video tutorial on this filter (in French), which can be seen on his Youtube channel (follow the link below)



Note that the construction of the function f(z) depends only on the random seed used to generate the expression tree. The figure below illustrates some image renderings obtained with different values of random seeds.




As the iterations proceed, more and more complex geometric patterns are formed in the image, as shown in the animation below.


By increasing the number of iterations and drawing only the contours of the subdivided triangles, we can generate truly remarkable fractal geometric patterns, given the apparent simplicity of the triangle subdivision algorithm, as illustrated in the following figure. This will give you ideas for tiling your kitchen or bathroom!




In G'MIC-Qt, the filter allows the user to enter a custom mathematical formula to define the function f(z) and its derivatives. It also offers options to colorize the fractal in a personalized way and to easily navigate in the fractal space thanks to its simple but efficient integrated browser. Finally, it extends the principle of calculation to other methods of resolution (Secant method and Householder method). And as always with G'MIC, it is also possible to script this filter from the command line (with the gmic command), to generate for example this type of animations:




In addition to the colors of the trunk and the leaves, the size of the branches and the probability of the branches splitting into several pieces can be adjusted. This allows generation of trees with quite different geometry and looks, as you can see on the figure below.


The following video illustrates the use of this filter in GIMP (click on the image below to access the video).



In the example above, the light rays stop when they meet the pixels of the G'MIC logo. But it is also possible to adapt the filter parameters so that the rays pass through the shape. This is done in the example below, where the user-defined mask corresponds to the glass surface of a living room window.


With this filter, it is easy to add custom lighting effects to photographs, as shown in the following example. All you have to do is define the mask of the shape guiding the light rays, and the filter does the rest!


The following video tutorial shows how to use this filter in GIMP:



The trick is to print the first image on a blank page and the second on a transparent sheet of the same size. By positioning the transparent sheet on the page in an appropriate way, and by moving it laterally, the animation becomes visible. Indeed, the vertical bars printed on the transparent layer allow only one image of the animation to pass through at a time, and these images follow one another as the transparent sheet moves:


Note that although the number of frames of the animation is indeed limited (in general, less than a dozen, there is obviously no possibility of converting a two-hour film!), there is no reason to limit oneself to grayscale images. The following two figures illustrate, for example, the rendering of a color animation with this filter.



"Your examples are nice, but what does it look like in real life?". Well, the feeling of animation is really present, as you can appreciate in the video below :

To sump up, it's an inexpensive and fun way to create personalized animation materials. Kids, not yet jaded by the Youtube application on their smartphone, will love it!

The set of filters we have just described is only a subset of all the new filters added in the last two years. In particular, there are many "community" filters that sometimes remain in the shadows for some time before being added to the main list of filters (e.g., filters in the Testing / category). To make it easier to keep track of these updates, G'MIC-Qt has added the About / What's New? function, which lists the recent additions and deletions of filters in the plugin. This list is tailored to the user, as it is based on the user's previous list.


As a reminder, the Update Filters button below the filter drop-down list will allow the plugin to check for new filters and download them if available.

This overview of the new G'MIC-Qt filters is now complete. But more improvements to the plugin are waiting for us!



Thanks to Nicholas Hayes, the author of the Paint.NET plugin, this 8bf version of G'MIC-Qt has been released. This is good news for all users of these proprietary software, who will be able to benefit from a new plugin well supplied with many image processing filters, with the possibility of modifying and improving them, and all this without paying anything (what a change! ☺).



The plugin has also seen its graphical interface improved, with in particular:





This button will delight all scripters: by pressing it, we get the equivalent G'MIC command to be invoked in the terminal for the gmic CLI tool, in order to apply the currently selected filter with its current parameter values. For example, here is what the clipboard contains after pressing this button in the example above:
This string can then be pasted back into a command line call to gmic, to get exactly the same effect on other images, or to integrate it into a more complex processing pipeline:


This is a really handy feature if you write your own processing scripts. You can take advantage of the convenience of the plugin's graphical interface (especially its preview window) to adjust the parameters of a filter, but also the power of the command line to apply batch processing to several dozen or hundreds of image files.

This description of the new features visible in the G'MIC-Qt plugin being completed, let's talk about the other improvements brought by this version 3.0. A lot of interesting things have been added to the core of the G'MIC project, namely its interpreter and its associated image processing library CImg (which is also updated to version 3.0). These new features are certainly less visible, but just as important since they have a potential impact on all the interfaces of the project!



This is of course only the beginning, but we hope to generalize the use of nn_lib to develop new interesting filters in the near future. More technical details about the nn_lib library can be found in the dedicated article on the G'MIC forum.





Moreover the G'MIC interpreter user has now the possibility to define and name their own color palettes which will be used in the map or index commands, used to apply these palettes on color images. These palettes are generally useful for color visualization of scalar image data. For example, the colorization of a Mandelbrot fractal can be done very simply in G'MIC with the following command line:




Secondly, it is now possible to export a 3D mesh object built in G'MIC (typically with a procedural generation) as an .obj file in Wavefront format, a simple ASCII format that most 3D modelers can also read. Here, illustrated below, is an example of importing a 3D object generated recursively in G'MIC and re-imported into Blender, using this Wavefront file format.



will show you this:



The field of digital image processing is essentially based on mathematical models of images and algorithms that we want to apply. G'MIC has its own mathematical expression evaluator, and this one has also seen many improvements recently. To summarize, more than 40 new mathematical functions have been implemented in this evaluator in the last two years, in order to improve its calculation capacities:
store(), date(), begin_t(), end_t(), merge(), f2ui(), ui2f(), ccos(), csin(), ctan(), ccosh(), csinh(), ctanh(), lerp(), maxabs(), minabs(), argmaxabs(), argminabs(), da_size(), da_insert(), da_push(), da_remove(), da_pop(), da_back(), isvarname(), resize(), fill(), repeat(), set(), deg2rad(), rad2deg(), swap(), vargkth(), vargmin(), vargmax(), vargminabs(), vargmaxabs(), vavg(), vkth(), vmin(), vmax(), vminabs(), vmaxabs(), vmed(), vprod(), vstd(), vsum(), vvar(), string().

More and more complex filters of G'MIC benefit from them. A lot of optimization work has also been done to make this evaluator more and more efficient, in particular by enabling parallel computing when possible. Rather than going into technical details about the implementation and optimization of this module, I demonstrate an example here that shows how this evaluator can be used in a G'MIC pipeline. Here, we want to produce a Buddhabrot image, a particular representation of the Mandelbrot set. Below on the left, you can see the content of the buddhabrot.gmic file defining the buddhabrot command, which produces the image on the right. The G'MIC mathematical expression evaluator is invoked between lines 9 and 27, which represents then a single math expression. As we can see, this evaluator is able to compile and execute math expressions which are in fact real programs!


The mathematical evaluator is an essential part of the G'MIC framework and will probably continue to evolve in future releases. Are we done with this article? Not yet! Because working on the G'MIC project obviously means spending a lot of evenings doing research, programming and testing new features, but an important part of the time spent also concerns the management of side items, namely: answering questions on forums or github forges, finding funding to make the project move forward faster, or communicating around the project (for example by writing this kind of articles ☺ !). And on these different points, notable things have happened, since the end of 2019:



In both cases, these financings were fruitful and enabled the opening of the G'MIC project to other horizons (not to mention the pleasure of interacting with these two talented engineers!).


Virtual Artist is an interesting way to support scientific interaction with the general public. It was used, for example, during the FÉNO 2021 (Festival of Excellence in Normandy), on the CNRS stand.






Since 2008, G'MIC has been an active project, even if we never defined a very precise roadmap. The development of new functionalities is done according to our free time and opportunities, as well as the research activities in image processing that we carry out in parallel in our research team at the GREYC laboratory. We only hope that this time will not ultimately end up vanishing! Being a researcher or a teacher-researcher sometimes seems to be incompatible with an open source software development activity, which is a very demanding job and at the same time not necessarily very well valued in a career evaluation. It can therefore be tempting to focus on more "profitable" activities from a professional point of view.

On a personal note, I would like to thank the GREYC laboratory) (notably the management, the management service and the DDA service), the CNRS Normandy Delegation) (notably the Valorisation service, the Communication service and the HR service), the INS2I Institute) of the CNRS, the LILA association, my colleague Sébastien Fourey (co-lead developer of the project), the members of the PIXLS.US association (which hosts our discussion forum) as well as all the contributors to the project, the users giving kind, useful, or even financial feedbacks. All these people who believed in the project at a given time and in the ability of its developers gave us their support in one way or another. This allowed us to never give up the development of G'MIC, since 2008. May they be truly thanked! To properly bring this report to a close, and to satisfy the readers who are still awake and remain hungry, here is a small selection of videos to occupy the long winter evenings that will arrive soon:

---

---