Multimode fiber (MMF) imaging encodes object information into complex speckle patterns, making high-fidelity reconstruction challenging. We propose a timing-consistent speckle-conditioned diffusion framework that aligns physical speckle propagation steps with diffusion time steps, guiding speckle reconstruction. A high-fidelity MMF simulator employing Gerchberg–Saxton preprocessing and high-order MM-GNLSE numerical integration generates physically consistent speckle sequences, which are embedded as conditional priors into a denoising diffusion probabilistic model. Experiments on a 10,310-group MNIST-MMF dataset demonstrate superior reconstruction performance, achieving a maximum correlation coefficient of 0.98 and consistent improvements over ResUNet and Pix2Pix across MAE, SSIM, and PSNR metrics. The proposed framework provides a new generative solution for robust MMF image reconstruction.