Improving generative modelling with Shortest Path Diffusion (ICML paper)

Diffusion models are currently the most popular algorithms for data generation, with applications in image synthesis, video generation, and molecule design. MediaTek Research has just published a paper showing that the performance of diffusion models is boosted by optimizing the diffusion process to minimize the path taken from the initial to the final data distribution. This contribution will improve the efficiency and quality of diffusion models, especially by reducing the resources necessary for generation and synthesis. The AI research group will present this work at the ICML conference in July.

• To learn more, the paper is available here
  
• An implementation of the algorithm is publicly available to use here