Draw Your Mind: Personalized Generation via Condition-Level Modeling in Text-to-Image Diffusion Models