Artificial intelligence (AI) models like Stable Diffusion can generate high-quality images based on words, but they often struggle to produce truly creative results when given vague terms like “creative.” Researchers at KAIST, working with NAVER AI Lab, have developed a way to make these models produce more creative images, such as unique chair designs, without needing extra training. This method improves the model’s ability to generate original and useful images.
The researchers focused on feature maps, which are internal data layers in AI models that help process and create images. They found that certain parts of the model, called shallow blocks, are key to making creative images. By converting these feature maps into a frequency domain, a mathematical space that separates data into low and high frequencies, the researchers could adjust the low-frequency parts to boost creativity. Low frequencies affect broad patterns, while high frequencies handle fine details. Amplifying low frequencies in shallow blocks helped create novel images without adding noise or strange colors. The researchers also created an algorithm, a set of instructions for the computer, to automatically choose the best amplification level for each block.
Improving image novelty and diversity
This approach allowed a pre-trained Stable Diffusion model to generate more creative images without additional data or training. The researchers tested their method and found it produced images that were more original than those from other models, while still being useful. They also solved a problem called mode collapse, where some models produce repetitive images, by increasing image variety in a faster model called SDXL-Turbo. Human evaluations confirmed that the images were noticeably more novel compared to older methods. The method works by manipulating the model’s existing capabilities, unlocking its hidden potential to create unique designs.
This research shows that AI models can generate creative images from text alone, which could inspire new ideas in fields like product design. The findings were presented at a conference on computer vision. The method is expected to make AI more practical for creative tasks, offering a simple way to enhance existing models without complex changes.