Learning From Matisse
LEARNING FROM MATISSE
Artists learn from past artists that set the groundwork of artistic ideas before them. In this series of still-life paintings, I am learning from 20th century painters. I am using digital collage and AI tools to automate the process of learning from many artists simultaneously and quantitatively defining the moment I am done learning having my own style. After using classifiers and GANs for over a year now, I think of them as automating the process of curating and applying a style respectively – essentially some of the processes an artist does as they are developing a painting or painting series.
Recently, I worked with a painting mentor who suggested I look at old masters artwork. By spending time with the masters and copying their paintings, she suggested I could learn painting techniques that can then be applied to my own art practice. I had an idea of how to automate this common practice of the painter’s learning process as well as create work that is not direct copies.
I wrote Python code that generate still life compositions by randomly placing colorful background shapes, vases, fruits and flowers. It’s actually the same code as my Machine Learning Abstracts, but rather than placing painterly shapes to create a geometric abstract, it is placing objects with more strict rules. For instance flowers must be placed directly above a vase.
Figure 1 – 3 sample randomly generated still life images generated from python code I wrote. Vases, containers and tables are collaged together. Some of the vases are objects in my studio, while others (like the one in the center image with blue patterning) was taken from a Matisse in his studio art book.
These generated compositions became my dataset to be processed with a cycleGAN to make them painterly looking. My painting dataset included my favorite painters: Cezanne, Matisse and Canadian artist Joseph Plaskett. Now with thousands of generated digital painting images, I used a TensorFlow Classifier, trained on the painting dataset to sort the images. With the images that are rated the most confident to be paintings, I sit down and physically paint them.
Figure 2 – 3 sample still lifes that have been processed by the cycleGAN to be more painterly.
Now with thousands of generated digital painting images, I used a TensorFlow Classifier, trained on the painting dataset to sort the images. With the images that are rated the most confident to be similar to the original painting dataset, I sit down and physically paint them by hand:
Figure 3 – A sample of 3 hand painted still lifes I created from working with the python code + cycleGAN + classifier (9″ x 7″ acrylic on paper, 2019)
I currently have painted over forty-five small still life paintings. I am inspired by the colors and new ideas generated from working with the code. I can start my studio time with literally hundreds of thousands of curated, compositions ready to paint. With consistent painting I noticed the confidence in my brush strokes increasing as well as how I address repetitive tasks. This learning and consciousness as I paint is extremely valuable when also working on automating the painting process with a robotic arm.
Figure 4 – thrity five and counting hand painted still lifes inspired by the dataset. The more I paint the more I retrain the algorithm with my own dataset. You can purchase these paintings here.
Finally here is a sneak peak at the latest status of the cycleGAN only trained on my images. It is intriguing as an artist to see how the algoirithm interprets a dataset I know very well.
Figure 5 – The cycleGAN trained only on the paintings I have done myself so far. Although there is not enough images yet to have the style really look like my painterly style, I love the patterns and shapes the GAN is coming up with. I see the blobs of color the same as how I love to layer discreet colors of paint.