The paper introduces FactorVAE, a method for unsupervised learning of disentangled representations from data generated by independent factors of variation. FactorVAE addresses the trade-off between disentanglement and reconstruction quality, which is a common issue with β-VAE. By augmenting the VAE objective with a penalty that encourages the marginal distribution of representations to be factorial, FactorVAE achieves better disentanglement scores than β-VAE while maintaining similar reconstruction quality. The paper also proposes a new disentanglement metric that is more robust and conceptually simpler than the metric used in Higgins et al. (2016), avoiding its limitations. Experimental results on various datasets demonstrate the effectiveness of FactorVAE in achieving better disentanglement without degrading reconstruction quality.The paper introduces FactorVAE, a method for unsupervised learning of disentangled representations from data generated by independent factors of variation. FactorVAE addresses the trade-off between disentanglement and reconstruction quality, which is a common issue with β-VAE. By augmenting the VAE objective with a penalty that encourages the marginal distribution of representations to be factorial, FactorVAE achieves better disentanglement scores than β-VAE while maintaining similar reconstruction quality. The paper also proposes a new disentanglement metric that is more robust and conceptually simpler than the metric used in Higgins et al. (2016), avoiding its limitations. Experimental results on various datasets demonstrate the effectiveness of FactorVAE in achieving better disentanglement without degrading reconstruction quality.