A possible future for image generation
(This paper was originally presented at the Proceedings of the Third International Conference on Generative Art, December 2000, Milan Polytecnic, Italy, and is published in the proceedings of the same conference.)
Years of trying to create an "Image Idea Generator" program have convinced me that the perfect solution would be to have an artificial artistic person, a design slave. This paper describes how I came to that conclusion, realistic alternatives, and briefly, how it could possibly happen.
1. The history of Repligator and Gliftic
In 1996 I had the idea of creating an "image idea generator". I wanted something which would create images out of nothing, but guided by the user. The biggest conceptual problem I had was "out of nothing". What does that mean? So I put aside that problem and forced the user to give the program a starting image.
This program eventually turned into Repligator, commercially described as an "easy to use graphical effects program", but actually, to my mind, is an "Image Idea Generator". The first release was in October 1997. In December 1998 I described Repligator V4  and how I thought it could be developed away from simply being an effects program.
In July 1999 Repligator V4 won the Shareware Industry Awards Foundation prize for "Best Graphics Program of 1999". Prize winners are never told why they won, but I am sure that it was because of two things:
1) Ease of use
2) Ease of experimentation
"Ease of experimentation" means that Repligator does in fact come up with new graphics ideas. Once you have input your original image you can generate new versions of that image simply by pushing a single key. Repligator is currently at version 7, but, apart from adding many new effects and a some new features, is basically the same program as version 4.
Following on from the ideas in  I started to develop Gliftic, which is closer to my original thoughts of an image idea generator which "starts from nothing". The Gliftic model of images is that they are composed of three components:
1. Layout or form, for example the outline of a mandala is a form.
2. Color scheme, for example colors selected from autumn leaves from an oak tree.
3. Interpretation, for example Van Gogh would paint a mandala with oak tree colors in a different way to Andy Warhol.
There could be a Van Gogh interpretation and an Andy Warhol interpretation.
Further I wanted to be able to genetically breed images, for example crossing two layouts to produce a child layout. And the same with interpretations and color schemes. If I could achieve this then the program would be very powerful.
1.2 Getting to Gliftic
Programming has an amazing way of crystalising ideas. If you want to put an idea into practice via a computer program you really have to understand the idea not only vaguely and globally, but just as importantly, in detail. You have to make hard design decisions, there can be no fuzziness. So implementing my ideas for Gliftic turned out to be a considerable challenge. I soon found out that the hardest thing to do would be the breeding of forms. What are the "genes" of a form? What are the genes of a circle, say, and how do they compare to the genes of the outline of the UK?
I wanted the genotype representation (inside the computer program's data) to be directly linked to the phenotype representation (on the computer screen). This seemed to be the best way of making sure that bred-forms would bare some visual relationship to their parents. I also wanted symmetry to be preserved. For example if two symmetrical objects were bred then their children should be symmetrical.
I decided to represent shapes as simply closed polygonal shapes, and the "genes" of these shapes were simply the list of points defining the polygon. Thus a circle would have to be represented by a regular polygon of, say, 100 sides. The outline of the UK could easily be represented as a list of points every 10 Kilometers along the coast line. Now for the important question: what do you get when you cross a circle with the outline of the UK? I tried various ways of combining the "genes" (i.e. coordinates) of the shapes, but none of them really ended up producing interesting shapes. And of the methods I used, many of them, applied over several "generations" simply resulted in amorphous blobs, with no distinct family characteristics. Or rather maybe I should say that no single method of breeding shapes gave decent results for all forms.
Figure 1 shows an example of breeding a mandala with 6 regular polygons:
Fig 1a Mandala Form
Fig 1b 6 regular polygons
Fig 1c Mandala bred with Polygons
I did not try out all my ideas, and maybe in the future I will return to the problem, but it was clear to me that it is a non-trivial problem.
And if the breeding of shapes is a non-trivial problem, then what about the breeding of interpretations?
I have temporarily abandoned the genetic (breeding) model of generating designs but retained the idea of the three components (form, color scheme, interpretation).
1.3 Gliftic today
Gliftic Version 1.0 was released in May 2000. It allows the user to change a form, a color scheme and an interpretation. The user can experiment with combining different components together and can thus home in on an personally pleasing image. Just as in Repligator, pushing the F7 key make the program choose all the options.
Unlike Repligator however the user can also easily experiment with the form (only) by pushing F4, the color scheme (only) by pushing F5 and the interpretation (only) by pushing F6.
Here are show some example images created by Gliftic.
Fig 2: A mandala interpreted with arabesques
Figure 3 Trellis interpreted with "graphic ivy"
Figure 4 Regular dots interpreted as "sparks"
1.4 Forms in Gliftic V1
Forms are simply collections of graphics primitives (points, lines, ellipses and polygons). The program generates these collections according to the user's instructions.
Currently the forms are:
Mandala, Regular Polygon, Random Dots, Random Sticks, Random Shapes, Grid Of Polygons, Trellis, Flying Leap, Sticks And Waves, Spoked Wheel, Biological Growth, Chequer Squares, Regular Dots, Single Line, Paisley, Random Circles, Chevrons.
1.5 Color Schemes in Gliftic V1
When combining a form with an interpretation (described later) the program needs to know what colors it can use. The range of colors is called a color scheme. Gliftic has three color scheme types:
1.6 Interpretations in Gliftic V1
Interpretation in Gliftic is best decribed with a few examples. A pure geometric line could be interpreted as:
1) the branch of a tree
2) a long thin arabesque
3) a sequence of disks
4) a chain
5) a row of diamonds.
An pure geometric ellipse could be interpreted as
1) a lake
2) a planet
3) an eye.
Actual real current interpretations:
Standard, Circles, Flying Leap, Graphic Ivy, Diamond Bar, Sparkz, Ess Disk, Ribbons, George Haite, Arabesque, ZigZag.
1.7 Applications of Gliftic
Currently Gliftic is mostly used for creating WEB graphics, often backgrounds as it has an option to enable "tiling" of the generated images. There is also a possibility that it will be used in the custom textile business sometime within the next year or two. The real application of Gliftic is that of generating new graphics ideas, and I suspect that, like Repligator, many users will only understand this later.
2. The future of Gliftic, 3 possibilties
Completing Gliftic V1 gave me the experience to understand what problems and opportunities there will be in future development of the program. Here I divide my many ideas into three oversimplified possibilities, and the real result may be a mix of two or all three of them.
2.1 Continue the current development "linearly"
Gliftic could grow simply by the addition of more forms and interpretations. In fact I am sure that initially it will grow like this. However this limits the possibilities to what is inside the program itself. These limits can be mitigated by allowing the user to add forms (as vector files). The user can already add color schemes (as images). The biggest problem with leaving the program in its current state is that there is no easy way to add interpretations.
2.2 Allow the artist to program Gliftic
It would be interesting to add a language to Gliftic which allows the user to program his own form generators and interpreters. In this way Gliftic becomes a "platform" for the development of dynamic graphics styles by the artist. The advantage of not having to deal with the complexities of Windows programming could attract the more adventurous artists and designers.
The choice of programming language of course needs to take into account the fact that the "programmer" is probably not be an expert computer scientist. I have seen how LISP (an not exactly easy artificial intelligence language) has become very popular among non programming users of AutoCAD. If, to complete a job which you do manually and repeatedly, you can write a LISP macro of only 5 lines, then you may be tempted to learn enough LISP to write those 5 lines.
Imagine also the ability to publish (and/or sell) "style generators". An artist could develop a particular interpretation function, it creates images of a given character which others find appealing. The interpretation (which runs inside Gliftic as a routine) could be offered to interior designers (for example) to unify carpets, wallpaper, furniture coverings for single projects.
As Adrian Ward  says on his WEB site: "Programming is no less an artform than painting is a technical process." Learning a computer language to create a single image is overkill and impractical. Learning a computer language to create your own artistic style which generates an infinite series of images in that style may well be attractive.
2.3 Add an artificial conciousness to Gliftic
This is a wild science fiction idea which comes into my head regularly. Gliftic manages to surprise the users with the images it makes, but, currently, is limited by what gets programmed into it and/or by pure chance. How about adding a real artificial conciousness to the program? Creating an intelligent artificial designer?
According to Igor Aleksander  conciousness is required for programs (computers) to really become usefully intelligent. Aleksander thinks that "the line has been drawn under the philosophical discussion of conciousness, and the way is open to sound scientific investigation". Without going into the details, and with great over-simplification, there are roughly two sorts of artificial intelligence:
1) Programmed intelligence, where, to all intents and purposes, the programmer is the "intelligence". The program may perform well (but often, in practice, doesn't) and any learning which is done is simply statistical and pre-programmed. There is no way that this type of program could become concious.
2) Neural network intelligence, where the programs are based roughly on a simple model of the brain, and the network learns how to do specific tasks. It is this sort of program which, according to Aleksander, could, in the future, become concious, and thus usefully intelligent.
What could the advantages of an artificial artist be?
1) There would be no need for programming. Presumbably the human artist would dialog with the artificial artist, directing its development.
2) The artificial artist could be used as an apprentice, doing the "drudge" work of art, which needs intelligence, but is, anyway, monotonous for the human artist.
3) The human artist imagines "concepts", the artificial artist makes them concrete.
4) An concious artificial artist may come up with ideas of its own.
Is this science fiction?
Arthur C. Clarke's 1st Law: "If a famous scientist says that something can be done, then he is in all probability correct. If a famous scientist says that something cannot be done, then he is in all probability wrong".
Arthur C Clarke's 2nd Law: "Only by trying to go beyond the current limits can you find out what the real limits are."
One of Bertrand Russell's 10 commandments: "Do not fear to be eccentric in opinion, for every opinion now accepted was once eccentric"
3. References and links
1. "From Ramon Llull to Image Idea Generation". Ransen, Owen. Proceedings of the 1998 Milan First International Conference on Generative Art.
2. "How To Build A Mind" Aleksander, Igor. Wiedenfeld and Nicolson, 1999
3. "How I Drew One of My Pictures: or, The Authorship of Generative Art" by Adrian Ward and Geof Cox. Proceedings of the 1999 Milan 2nd International Conference on Generative Art.
4. Repligator home page: http://www.repligator.com/