Art-Free Generative Models: Art Creation Without Graphic Art Knowledge
Hui Ren, Joanna Materzynska, Rohit Gandikota, David Bau, Antonio Torralba
2024-12-03
Summary
This paper explores how art can be created without needing extensive knowledge of graphic art by developing a model that generates images based solely on text descriptions.
What's the problem?
Creating art typically requires a deep understanding of artistic techniques and styles. Many existing models for generating art rely on large datasets filled with artistic content, which means they need prior knowledge about art to produce good results. This can limit who can create art using AI, as not everyone has that background.
What's the solution?
The researchers propose a new text-to-image generation model that does not use any art-related content during training. Instead, they introduce a method to learn an 'art adapter' using only a few examples of specific artistic styles. This allows the model to generate art that users perceive as comparable to pieces created by models trained on extensive art datasets. They also analyze how different types of data contribute to creating new artistic styles.
Why it matters?
This research is significant because it opens the door for more people to create art using AI, regardless of their background in graphic design or art history. By demonstrating that high-quality art can be generated without extensive prior knowledge, this work encourages creativity and innovation in using technology for artistic expression.
Abstract
We explore the question: "How much prior art knowledge is needed to create art?" To investigate this, we propose a text-to-image generation model trained without access to art-related content. We then introduce a simple yet effective method to learn an art adapter using only a few examples of selected artistic styles. Our experiments show that art generated using our method is perceived by users as comparable to art produced by models trained on large, art-rich datasets. Finally, through data attribution techniques, we illustrate how examples from both artistic and non-artistic datasets contributed to the creation of new artistic styles.