Inversion-Based Creativity Transfer with Diffusion Models

In this paper, we introduce the task of "Creativity Transfer". The artistic creativity within a painting is the means of expression, which includes not only the painting material, colors, and brushstrokes, but also the high-level attributes including semantic elements, object shape, etc. Previous arbitrary example-guided artistic image generation methods (e.g., style transfer) often fail to control shape changes or convey semantic elements. The pre-trained text-to-image synthesis diffusion probabilistic models have achieved remarkable quality, but they often require extensive textual descriptions to accurately portray attributes of a particular painting. We believe that the uniqueness of an artwork lies precisely in the fact that it cannot be adequately explained with normal language. Our key idea is to learn artistic creativity directly from a single painting and then guide the synthesis without providing complex textual descriptions. Specifically, we assume creativity as a learnable textual description of a painting. We propose an attention-based inversion method, which can efficiently and accurately learn the holistic and detailed information of an image, thus capturing the complete artistic creativity of a painting. We demonstrate the quality and efficiency of our method on numerous paintings of various artists and styles. Code and models are available at https://github.com/zyxElsa/creativity-transfer.