On Improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models