We propose TTA -Diffusion(Token Timestep Allocation), that stabilizes the control process of diffusion language models by dynamically allocating timesteps per token, addressing instability that disrupts fluency and control.