Enhancing Creativity in AI: A Breakthrough from KAIST

Photo 1. Professor Jaesik Choi, KAIST Kim Jaechul Graduate School of AI
Credit: Statistical Artificial Intelligence Lab @KAIST
View more

In the ever-evolving landscape of artificial intelligence, particularly in the realm of generative models, recent advancements have sparked exciting conversations about the intersection of technology and creativity. One notable development comes from Professor Jaesik Choi and his research team at KAIST Kim Jaechul Graduate School of AI, who have devised a method to enhance the creativity of text-based image generation models such as Stable Diffusion. This breakthrough promises to revolutionize how we leverage AI in creative fields, from product design to the arts.

The Challenge of Creativity in AI

Text-based image generation models, like those powered by Stable Diffusion, have gained prominence for their ability to produce high-quality images from natural language descriptions. Yet, a limitation persists: their creative output can sometimes feel formulaic or constrained. For instance, if presented with a prompt like "creative," the results can often miss the mark in terms of true inventiveness.

To address this challenge, Choi’s team, in collaboration with NAVER AI Lab, explored ways to invigorate the creativity inherent in these models without depending on additional training. Their approach leverages an innovative method of manipulating the internal feature maps of generative models, thus unlocking new creative potentials without overhauling the underlying system.

Technical Insights into Creative Amplification

The crux of the team’s research lies in the selective amplification of internal feature maps within AI models. Specifically, they identified that certain shallow blocks are vital for fostering creative output. Through their experiments, they found that amplifying frequency values in the high-frequency region could lead to noisy or fragmented results. Instead, focusing on low-frequency regions within those shallow blocks proved to effectively boost creative generation.

Building upon this insight, the researchers created an algorithm that autonomously determines the optimal amplification for each block in the generative model. This approach opens the door to generating more diverse and original images without necessitating new classification data or intensive training sessions.

Quantitative Validation and User Studies

To substantiate their findings, the research team employed various metrics to assess the innovative capabilities of their algorithm. They demonstrated that their technique results in images characterized by greater novelty compared to those produced by existing models, without significantly diminishing their practical utility.

Notably, the research team tackled a common issue in generative models known as mode collapse—where the model generates very similar outputs despite varying inputs. Employing their algorithm in the SDXL-Turbo model, which aims to accelerate image generation in the Stable Diffusion XL (SDXL), they observed a marked increase in image diversity and creativity.

Additionally, qualitative feedback from user studies corroborated their quantitative results, revealing that users found the images produced through their method significantly more novel while still retaining utility.

Voices from the Research Team

Jiyeon Han and Dahee Kwon, Ph.D. candidates at KAIST and co-first authors of the published research, expressed their enthusiasm for their findings: "This is the first methodology to enhance the creative generation of generative models without new training or fine-tuning. We have shown that the latent creativity within trained AI generative models can be enhanced through feature map manipulation."

They further emphasized the potential implications of their research, noting, "This research makes it easy to generate creative images using only text from existing trained models. It is expected to provide new inspiration in various fields, such as creative product design, and contribute to the practical and useful application of AI models in the creative ecosystem."

Publication and Academic Impact

The findings of this innovative research were presented at the International Conference on Computer Vision and Pattern Recognition (CVPR) on June 16, 2025. The paper, titled Enhancing Creative Generation on Stable Diffusion-based Models, can be found on arXiv.

The study received substantial support from several initiatives, including the KAIST-NAVER Ultra-creative AI Research Center and various government-funded projects aimed at advancing AI technology while adhering to ethical considerations.

This research represents a significant leap forward in the capabilities of AI in creative domains, allowing for greater flexibility and innovation in applications where imagination and originality are paramount.