This paper presents a deep learning framework for pricing and calibration of stochastic volatility models, with a focus on rough volatility models that are notoriously slow to calibrate due to their reliance on Monte Carlo methods. The authors propose a two-step approach: first, a deep neural network is trained offline to approximate the mapping from model parameters to the implied volatility surface (or option prices) using supervised learning on numerically generated data. Second, the trained network serves as a fast surrogate for online calibration, reducing calibration time from minutes or hours to milliseconds. The method is demonstrated on several models including the rough Bergomi, Bergomi, Heston, and SABR models, showing that the neural network approximation achieves accuracy comparable to the original numerical method (e.g., Monte Carlo) for out-of-sample parameters. Key innovations include representing the implied volatility surface as a collection of pixels for efficient learning and handling piecewise constant forward variance curves in the rough Bergomi model. Numerical experiments on synthetic and historical SPX data confirm that the approach enables fast, accurate, and robust calibration, effectively removing the computational bottleneck that previously limited the industrial applicability of rough volatility models. The authors also discuss potential extensions to model recognition and hedging.