Improving the Stability of Diffusion Models for Content Consistent Super-Resolution