ProCreate, Dont Reproduce! Propulsive Energy Diffusion for Creative Generation
Jack Lu, Ryan Teehan, Mengye Ren
2024-08-06
Summary
This paper presents ProCreate, a new method designed to enhance creativity and diversity in image generation models, while preventing them from simply copying training data.
What's the problem?
Many current image generation models struggle with producing diverse and creative outputs. They often end up reproducing the training data instead of generating unique images. This happens because the models rely too much on reference images, leading to repetitive results.
What's the solution?
ProCreate addresses this issue by actively pushing the generated images away from the reference images during the creation process. This method encourages the model to explore new ideas and styles instead of just mimicking what it has seen before. The authors also introduced a dataset called FSCG-8, which includes various categories of creative content, helping ProCreate achieve greater diversity and quality in its outputs.
Why it matters?
The significance of ProCreate lies in its ability to foster creativity in AI-generated images. By improving sample diversity and preventing data reproduction, this method can lead to more innovative and interesting visual content. This is important for fields like art, advertising, and entertainment, where unique visuals are essential.
Abstract
In this paper, we propose ProCreate, a simple and easy-to-implement method to improve sample diversity and creativity of diffusion-based image generative models and to prevent training data reproduction. ProCreate operates on a set of reference images and actively propels the generated image embedding away from the reference embeddings during the generation process. We propose FSCG-8 (Few-Shot Creative Generation 8), a few-shot creative generation dataset on eight different categories -- encompassing different concepts, styles, and settings -- in which ProCreate achieves the highest sample diversity and fidelity. Furthermore, we show that ProCreate is effective at preventing replicating training data in a large-scale evaluation using training text prompts. Code and FSCG-8 are available at https://github.com/Agentic-Learning-AI-Lab/procreate-diffusion-public. The project page is available at https://procreate-diffusion.github.io.