Yilun Xu and his team have developed a new way for neural networks to generate images using the physical process of electric field creation by charged particles. This method, Poisson Flow Models (PFGM), creates images of similar quality to diffusion-based methods but 10-20 times faster.
The team extended their model to include different dimensions, allowing researchers to fine-tune a neural network’s robustness and ease of training. Future work will focus on finding the best balance between these factors and exploring other physical processes that could form the basis for new generative models.