Medical image generation faces challenges due to data scarcity, high annotation costs, and the need for precise structural fidelity. While diffusion models excel at generating high-quality images, they suffer from slow sampling speeds. Conversely, autoregressive models, although efficient in sequence modeling, struggle with continuous visual signals. In this work, we propose a diffusion-to-autoregressive framework for lung CT image synthesis, where the autoregressive model learns the temporal denoising dynamics of a diffusion process. Specifically, we introduce a sequential diffusion tokenizer that discretizes intermediate diffusion latents into structured token sequences, enabling the AR transformer to perform next-step predictions on diffusion trajectories rather than raw images. This formulation preserves the coarse-to-fine generation behavior of diffusion models while substantially accelerating sampling. Additionally, we incorporate a multi-condition guidance mechanism that utilizes segmentation masks and anatomical structures to enhance generation controllability. Experiments on a low-dose lung CT dataset show that our method achieves outstanding performance in FID, IS, and SSIM metrics. Furthermore, a Visual Turing Test conducted with radiologists confirms the perceptual realism of the generated images. Our approach has potential for extension to multi-task and multi-modality medical imaging applications.