Benefit from its capability of learning high-dimensional compact representation from raw data, the auto-encoders are widely used in various tasks of data compression. In particular, for deep image compression, auto-encoders generally take the responsibility of mapping original images to the latent representation to be coded. In this paper, we propose a new framework for deep image compression by devising a loss function for latent optimization, and adopting the differentiable approximation of quantization. In our experiments, both subjective and objective results can confirm the effectiveness of our contributions.