Optimizing Denoising Diffusion Probabilistic Models (DDPM) with Lightweight Architectures: Applications of SqueezeNet and MobileNet

In the field of image generation technology, enhancing Denoising Diffusion Probabilistic Models (DDPM) to boost generation efficiency and decrease computational demands is especially crucial. This study investigates the optimization of DDPM by integrating lightweight neural network architectures, specifically SqueezeNet and MobileNet, into the U-Net framework. The goal is to enhance image generation efficiency while minimizing computational requirements, making these models suitable for real-time applications in resource-constrained environments. Experimental results show that both models effectively reduce Mean Square Error (MSE) loss, with MobileNet achieving a final average loss of 0.0327, significantly lower than SqueezeNet's 0.0605. MobileNet also demonstrated superior GPU utilization at 9%, compared to SqueezeNet's 15%. However, the Denoising Diffusion Implicit Model (DDIM) outperformed both lightweight models, achieving a final average loss of 0.0001 due to its advanced strategies like skip sampling. These findings highlight the potential of lightweight architectures in diffusion models and suggest that future research should focus on enhancing these models by incorporating advanced techniques to improve image quality. This study contributes to the discourse on efficient image generation techniques, paving the way for broader applications of diffusion models.