Rothko Generator Project
authors: Maya Kalogera and CSDVU
Rothko generator - # untitled till infinity is software/installation which deals with the proliferation, overload and recycling of electronic pictorial waste, incorporating it into CG images based on Rothko's work.Artist's Statement
As internet is full of boring banners, pornography, dull images, etc., I decided to make an automatic image generator which will make "Rothko's paintings" where pixels will be substituted with the images from the internet. I like his paintings particularly and it is a challenge to make computer generated paintings visually based on Rothko's work from late 50's and early 60's, which will achieve high level of "authenticity" due to its use of texture, hue alterations and oblong shape modifications so characteristic for his work at that period.
Beside that intention, this project lies within my interest in collecting datas from the Internet and using that informations to create a completely different visual outcome. In this project, those two streams, infinitely CG images and collected profane net pictures, amalgamate themselves in very appropriate way.
Rothko once said: 'It is a widely accepted notion among painters that it does not matter what one paints as long as it is well painted. This is the essence of academicism. There is no such thing as good painting about nothing.'
For Rothko, color is "merely an instrument." Rothko demanded his paintings be organic, art that "lives and breathes," an object among others. He noticed: "The fact that people break down and cry when confronted with my pictures shows that I can communicate those basic human emotions... the people who weep before my pictures are having the same religious experience I had when painting them. And if you say you are moved only by their color relationships then you miss the point." His work concentrated on basic emotions, often filling the canvas with very few, but intense colours, using little immediately-apparent detail. The rarest sense for fine saturated color values, in from of us there aren't dead abstract surfaces, but the surfaces which breathe, emanate their inner glow. Rothko repeatedly protested: 'I'm not interested in color' and 'I'm not a colorist .' Color, he explained, was nothing more than an 'instrument "for expressing something larger: the all important 'subjects' of his pictures". Rothko said for his paintings that they are organic, that they breathe. At the micro- structure they indeed change their inner glow: spiritual and mystic is suggested by colors and their displaying order.
And rightly to that mystic feeling, it seemed to me as a great challenge to confront something so profane as are advertisements, porn pics, charts, banners, etc., and yet to camouflage it and leave that colored superficies to emanate.
Spatial installation of RGenerator / untitled till infinity :
Viewers enters among projections, they are surrounded with them from at least 3 sides. Colours are slowly changing, melting, generating new hues, transforming into new shapes. Viewer can sit, rest there, for no matter how long, the images will always be new ones, somehow more or less similar, but never ever the same.
That continous stream in infinity with significant rhythm changes (from leggero to moderato, with never the same pace), makes color significance more stronger, vivid, expressive. Like at the sacral place, energy arises by the lenght of time being spent amongs the paintings.
I hope that the viewer will feel exactly as in Dominique De Menil's saying about Rothko's paintings:"They embrace us without closing us. Their dark surfaces do not stop the gaze. A light surface is active - it stops the eye. But we can gaze right through these purplish browns, gaze into the infinite."
Rothko Generator Project is still in beta phase, meaning that still some improvements on souceOver matrix are on its way. But even though, the
final output has alredy the visual appearance which was intented to
For purposes of this project, 18 Rothko’s paintings were analysed and
treated by same hypothesis: as they were consisted of background and foreground surfaces. It wasn’t always applicable smoothly, but it considerably facilitated the analysis of shapes and colors. Other element which further alleviated the process was presuming that shapes in paintings are places on different co-ordinates on z-axe: smaller shapes have higher co-ordinate value, bigger forms have lower co-ordinate value. By this factor it was gained bigger maneuvering space for later manipulation and placement of new generated forms and for their use in sourceOver matrix (alpha blending).
One of the biggest obstacles at the begining of the process was how to gain "realistic" shape modelling similar to forms on Rothko's paintings. This was solved with Kok F.Lai's Gsnake image processing library for edge definitions. Gsnake provides tools for contour modeling, extraction, detection and classification, based on generalized active contour model. It consists of a set of objects, suitable for use in the area of feature extraction, character recognition, motion analysis and etc.
The closest technique which was the initial point for developing the algorhitms
in this project is photomosaic. The typical photomosaic algorithm searches from
a large database of images one picture that approximates a block of pixels in
the main image. Since the quality of the output depends on the size of the database, it turns out that the bottleneck in each such algorithm is the searching process.
In this project we used a technique to speed-up this critical phase using the Antipole Tree Data Structure. This improvement allows the use of larger databases without requiring much longer processing time. This is a new method invented by Gianpiero di Biasi to speed-up the search process. This technique is based on the Antipole Tree Data Structure and allows us to obtain impressive effects in an efficient manner. The Antipole Tree is suitable for searches over large record sets embedded into a metric space (X, d).
The Antipole Clustering of bounded radius is a top-down procedure that starts with a given finite set of points X in a metric space. The first check is if there exists a pair of points in X such that their distance is longer than the radius. If this is the case, the set is partitioned by assigning each point of the splitting subset to the closest endpoint of the pair (A, B).
Otherwise the splitting is not performed and the given subset is itself a cluster. Once the data structure is built a suitable nearest neighbor algorithm can be designed.
The search, starting from the root, proceeds by following the path in the tree, which guarantees to find the nearest cluster centroid pruning the impossible branches. A backtracking search explores the remaining branches of the tree
to assure a correct answer.
RG algorithms can be divided into two phases: database acquisition and RGB value recognizer/pixel replacer process.
The acquisition of the database of images is very simple: each image of the database is partitioned into 9 equal rectangles arranged in a 3x3 grid and compute the RGB mean values for each rectangle. This leads to a vector x composed by 27 components (three RGB components for each rectangle). x is the feature vector of the image in the data structure. When all the images in the database have their own feature vector the Antipole clustering can be performed.
The painting creation explained in few steps: first the input image is subdivided into a regular grid, then each cell of the grid into another 3x3 sub-grid. Second the RGB mean values for each sub-cell of the sub-grid are computed. This leads to a vector x composed by 27 components (three RGB components for each sub-cell). x is the feature vector of the cell. After performing the best matching we resize the selected tile to fit and paint it over the cell with RenderBot.
The final image is created when the simulation is finished by having each RenderBot execute its painting function.
RenderBot classes differ in their physical behavior as well as their way of painting so that different styles can be created in a very flexible way. So they provide an unified approach for stroke based rendering.
Final visual output is a computed motion map between two generated images, and warped one into the another through a smooth animation.
Finally it is pointed out that the algorithm for RG output is completely automated and requires no human hints. In particular there is no need to manually select the directional guidelines in the input image or at other stage of painting production. This is a heuristically based, efficient and direct geometric construction scheme that leads to generated "Rothko paintings" with a good visual impact.