Chinese landscape paintings ‘drawn’ by AI platform Sapgan are hailed for their creativity, message, and artistic values that equal to true artists.
In 2019, Nvidia had introduced the Artificial Intelligence (AI) ‘artist’ GauGAN, which allows users to create landscape paintings by just drawing simple strokes. GauGAN utilizes a deep learning technology called Generative Adversarial Network, or GAN, often used by AI insiders in creating images. Upon its introduction, GauGAN had attracted so much attention that the platform’s official website crashed due to overwhelming user traffic.
However, most paintings created by AI are in Western art styles, from realism, postmodernity, or abstract. There is seldom any AI artist that assumes traditional Oriental arts. This had inspired Alice Xue, an undergraduate at the prestigious Princeton University, to develop an AI model called Sapgan (shortened from Sketch-And-Paint GAN) which can create astounding traditional Chinese landscape paintings. Her project was named the most excellent thesis at Princeton for the class of 2020.
Sapgan’s drawing process is much similar to human artists: it sketches, then paints. To let her AI follow the traditional drawing process of Chinese landscape paintings, which includes: inking, sketching, forming, decorating, painting… Alice had programed two phases for the AI, called SketchGAN and PaintGAN.
In particular, SketchGAN will collect high-resolutions images of traditional Chinese landscape paintings, then make according sketches and transferring the task to PaintGAN for completion.
According to a small scale survey over 242 people, 55 percent believe that Sapgan’s works look more ‘real’ than those done by human artists. Even Chinese people who are familiar with the genre, gets confused with these drawings.
Alice also did a survey between people with Chinese origins and Western people. Results showed that while only 49 percent of Chinese people can accurately identify between real paintings and those done by AI, 73.5 percent of Western people gave correct answers. This means that the more familiar people are with this traditional art style, the more they struggle to tell these drawings apart.
When asked to evaluate the ‘creativity’, ‘wisdom’, ‘aesthetics’ of these paintings, every participant give higher scores to those done using algorithms. The only element that people can recognize is the paintings’ ‘clarity’.
What’s surprising about this project is that its creator was never an AI major. She collected over 2,192 Chinese landscape paintings on her own to create a training database for her deep learning model, instead of coding algorithms that can collect by themselves. This may be the key element that lead to the more ‘natural’ look of SAPGAN’s paintings in comparison to other artworks.
According to Alice, the number of paintings in her training data is quite the disadvantage, however. She believes that the lack in both quantity and quality so far had limited the number of outputs for her model. SAPGAN’s paintings are not commercialized, but rather are uploaded to GitHub for communal use.
AI EraRelated posts: