In this work we present a diffusion transformer (DiT) framework with hybrid conditioning for machine learning based structural topology optimization. Traditional finite element based methods require repeat complex analysis that makes it impractical for real time design applications. Machine learning approaches have recently developed with the aim of accelerating topology optimization by learning direct mappings from problem definitions to optimal topologies. We propose a diffusion transformer-based generative model for two-dimensional structural topology optimization that leverages multiple modern conditioning methods to incorporate both local physics and global design constraints into the topology generation process. Our approach combines spatial conditioning via concatenated Von-Mises stress and strain energy fields with global conditioning through adaptive layer normalization informed by load parameters and prescribed volume fraction. Multiple DiT model scales and patch sizes are evaluated to study the trade-off between model inference speed and structural compliance accuracy. Our results demonstrate errors below 1% when compared to ground-truth solutions, accurate enforcement of volume constraints, and negligible presence of disconnected components. We also demonstrate that deterministic sampling with reduced diffusion steps enables faster inference while maintaining high quality structural designs.