Back to today’s briefing

Story perspectives

New Study Links Diffusion Models to Human Creativity

9/28/2025

50 12

Summary

Story summary

Researchers Kamb and Ganguli show that diffusion-model creativity arises from the denoising step, emphasizing locality and equivariance.
Diffusion models blend training-image elements to produce new, coherent outputs rather than memorize them, and the study compares this process to human creativity.
Kamb's hypothesis led to the equivariant local score (ELS) machine, which matches outputs of trained diffusion models with 90% accuracy.

1 of 1

Story summary

Researchers Kamb and Ganguli show that diffusion-model creativity arises from the denoising step, emphasizing locality and equivariance.
Diffusion models blend training-image elements to produce new, coherent outputs rather than memorize them, and the study compares this process to human creativity.
Kamb's hypothesis led to the equivariant local score (ELS) machine, which matches outputs of trained diffusion models with 90% accuracy.