FPGA-Based Diffusion Model for Image Super-Resolution

In recent years, diffusion models have achieved state-of-the-art performance in image super-resolution (SR). However, efficiently deploying these models on FPGAs remains a significant challenge due to the excessive latency and memory demands inherent in iterative generation. To address these issues, this paper design an image super-resolution reconstruction system based on diffusion models. With a significantly reduced number of iterations, each module is custom-designed, and the storage scheme is optimized through buffer partitioning to enable efficient parallel computation. The proposed system is implemented and evaluated on a Xilinx XCZU15EG FPGA platform. Experimental results show that the FPGA system achieves PSNR and SSIM comparable to those of a GPU, while improving energy efficiency by approximately 1.68 times compared to the NVIDIA GeForce RTX 4060.