Diffusion models have been successful in image and video generation, and now, they can generate high-performing neural network parameters. By using an autoencoder and a latent diffusion model, our method extracts representations of trained network parameters, synthesizes them from noise, and generates new representations that outperform or match trained networks. Surprisingly, the generated models perform differently from the trained ones. This approach shows promise for creating models with improved performance at minimal cost, sparking further exploration of diffusion models’ versatile applications in neural networks.
https://arxiv.org/abs/2402.13144