This paper introduces a novel approach for generative modeling of limit order book (LOB) data by converting LOB states into a structured image format and applying diffusion models with inpainting. The method leverages spatio-temporal inductive biases through a convolutional UNet backbone, enabling parallel generation of long sequences and avoiding error accumulation common in autoregressive models. The authors contribute to the LOB-Bench benchmark to allow fair comparison between models using Level-2 (order book only) and Level-3 (including message data) inputs. Despite using only Level-2 data, the proposed model achieves state-of-the-art performance on GOOG stock under Wasserstein loss and competitive results under L1 loss, outperforming existing methods like LOBS5 on certain metrics. However, performance on INTC (a large-tick stock) is weaker due to sparse signal, leading to mode collapse. Ablation studies show that inference speed can be significantly improved with minimal performance loss by reducing sampling steps, highlighting the method's efficiency. The work demonstrates that diffusion-based image generation is a promising paradigm for financial LOB simulation, particularly when global distributional similarity is prioritized over local fidelity. Future directions include latent-space diffusion, conditional generation, and improvements for large-tick stocks.