AI has learned how to convincingly paint in the style of masters like Van Gogh and Vermeer.
Dinesh Manocha at the University of Maryland and colleagues trained an AI to replicate the styles of well-known paintings including the pointillist The Eiffel Tower by Georges Seurat, The Starry Night by Vincent Van Gogh, oil paintings by Johannes Vermeer, and Chinese ink landscapes by Ming dynasty painter Shen Zhou.
For each style, the AI was trained with three to 10 reference paintings via a method known as reinforcement learning. It uses trial and error so the AI learns from its own actions and experiences.
Advertisement
For each reference painting, the algorithm learned to identify the properties of brush strokes, including their position, density, size and colour, and what order to paint them in.
The training process for one style took about six hours, as the AI learned to replicate the reference paintings. It learned that the more closely its painting resembled the original, the better it was performing.
The team then presented the AI with a new image 鈥� for example, a photograph of the Eiffel Tower at night, or a vase of sunflowers 鈥� and tasked it with reproducing it as a painting, in the style of Vermeer, for example.
The algorithm then generated brush strokes step-by-step, taking about five minutes to complete a single painting.
Existing filters can turn photographs into painting-like images, but because this algorithm paints thousands of strokes sequentially, the end result is more realistic.
The team includes researchers at Adobe Research, so the AI may eventually make it into a feature in Adobe鈥檚 graphics editing software Photoshop.
AI has been used to generate paintings and images, and a portrait created by an algorithm sold for over USD $400,000 at auction in October last year.
鈥淚n terms of painting techniques, the capability of an AI-based system will be close to or exceeds that of a human,鈥� says Manocha. But algorithms are still no match for human creativity, he believes.
搁别蹿别谤别苍肠别:听arXiv,听
