An Efficient Diffusion-Based Framework for Accelerated Image Generation Using Fast-DDPM

Abstract: Denoising Diffusion Probabilistic Models (DDPMs) have emerged as a leading class of generative models for image synthesis, yet their reliance on long Markov chains of up to one-thousand-time steps makes them impractical for many medical imaging applications, where training a single model can take days and generating one image volume can take minutes to hours. This paper presents a comprehensive analysis and implementation framework built around Fast- DDPM, an efficient diffusion-based approach that aligns the training and sampling procedures of DDPMs so that both stages operate over only ten time steps rather than the conventional one thousand. Two complementary noise-scheduling strategies, a uniform time-step scheduler and a non-uniform time-step scheduler, are described in detail, along with the modified forward and reverse processes that allow the denoising network to be trained efficiently without sacrificing sample fidelity. The framework is evaluated on three representative medical image-to-image generation tasks: multi- image super-resolution of prostate MRI, denoising of low-dose lung CT scans, and cross-modality translation of brain MRI. Across all three tasks, the accelerated framework matches or exceeds the image quality of the full DDPM baseline while reducing training time to roughly one-fifth and sampling time to roughly one-hundredth of the original cost. These results indicate that carefully aligning the number of training and sampling steps, rather than relying solely on post hoc fast samplers or expensive distillation procedures, is sufficient to close the efficiency gap that has limited the clinical adoption of diffusion models. We further discuss the implications of this efficiency gain for real-time clinical workflows, the trade-offs revealed by an ablation over the number of time steps, and directions for extending the framework to three- dimensional and four-dimensional medical imaging data.

Keywords: Diffusion models, Denoising Diffusion Probabilistic Models, medical image generation, image super- resolution, image denoising, image-to-image translation, fast sampling, computational efficiency.