A new study, highlighted in Quanta Magazine, reveals the mechanism of "creativity" in AI diffusion models like Midjourney. Scientists have concluded that creativity is not a magical property but a deterministic byproduct of the models architecture itself. The image generation process begins with random noise. At each of the many steps, the model slightly "denoises" the image, bringing it closer to the text prompt. The researchers found that due to architectural constraints (specifically "locality," where the model focuses on small patches rather than the whole picture at once), the AI is forced to "improvise," assembling the final image from individual fragments like a mosaic. It is this sequence of local decisions that gives rise to new, original compositions, rather than a simple averaging of images from the training data. This discovery helps to demystify the creative abilities of AI and opens up avenues for creating more controllable generative tools.
Researchers Explain AI "Creativity" Through the Denoising Process
