Lin
Generating Surrealism Art
This is an independent research project on generating surrealism art in Imperial College London.
Image2Art: Generating Your Surrealism
Hexun Lin, Thomas Lancaster (Supervisor)
Imperial College London
This was a 9-month long independent research project on generating surrealist art, divided into two periods. The first period focused on doing some research on methods of generating art and testing six methods using a dataset collected by myself in preparation for the second phase. The second period aimed to investigate an innovative method to generate surrealist art, and based on the results of the first period, the CLIP-Generator method was chosen as the base framework for an in-depth study. Proposed an innovative probability-based pipeline to convert images into surrealism art.
Period I:
Methods Review and Technical Tests for Generating Surrealist Art
This independent study is a semester-long project under the supervision of Thomas Lancaster, which is dedicated to doing a comprehensive survey of traditional and emerging modern art generation techniques, with a particular focus on the surrealism genre. Tested and summarized the advantages and disadvantages of various techniques, such as Style Transfer, Deep Dream, DCGAN, ACGAN, CycleGAN, BIGGAN, etc., and got different generative artworks. Also looking for ways to quantify and evaluate art creation to help people explore new ways to produce art.
After the experiments, it was found that each model has its own pros and cons, with CLIP-BIGGAN being the best-performing one. Besides, a survey was also done in these experiments to help evaluate different results of each method with 40 participants of different ages and backgrounds. There are five same questions for each image, and each question has 10 points. The artworks with the highest scores in human-like, creative, emotional, surreal, and popular degrees were mostly produced by Clip-BigGAN and Style Transfer. It showed that people cannot well distinguish the AI-generated surrealism art from human works, and also their attitudes towards the performance of each method.
Period II:
Image2Art: A Zero-Shot Method Based on Probability to Generating Surrealism Art
Surrealism, as a form of art, can express human intuition and subconsciousness fully. The uncertainty of machine learning greatly aids the generation of surrealist elements, which is a topic with great value and potential. This paper aimed to introduce an image-to-art pipeline that performed a surrealist reconstruction of the source domain by generating a target domain expected to be a different class but with some similarity in shape or color—for example, turning the cloudy sky in the image into a flying cloak. Our pipeline was implemented using the CLIP model twice, which produced the text prompts automatically with the maximum probability of labels (from ImageNet), to guide the generator to output surrealist art. We also added the similarity between the source image and the output to the generator’s loss function to ensure the structural similarity. Eventually, we showed that Image2Art is capable of creating fantastic artwork with surrealist elements.
