This study introduces a novel generative adversarial network(GAN)-based framework incorporating the Girsanov theorem to transform historical log-return series into risk-neutral paths, facilitating effective and arbitrage-free option pricing. While previous GAN applications to financial time series primarily focused on sequence generation using general-purpose architectures, this research comparatively evaluates GAN models, specifically TimeGAN, QuantGAN, and SigCWGAN, against conventional Monte Carlo simulations across various market conditions and option types. Empirical findings indicate that Girsanov-transformed models, particularly QuantGAN and SigCWGAN, outperform MC simulations in terms of pricing accuracy and distributional fidelity under stable market regimes. However, model performance notably deteriorates during periods of heightened volatility and extreme option moneyness levels, revealing the sensitivity of GAN-based approaches to abrupt market regime shifts. These results emphasize the critical importance of integrating volatility-aware modeling and adaptive retraining strategies in option pricing via deep generative models.