Atrial fibrillation (AF) is the most common sustained arrhythmia, and its management requires accurate characterization of atrial electrical activity. Electrocardiographic imaging (ECGI) and deep learning (DL) methods aiming at estimating electrograms (EGMs) noninvasively from body surface potentials (BSPMs) are promising, but progress is limited by the scarcity of paired BSPM–EGM datasets. Thus, we explore variational autoencoders (VAE) to generate synthetic EGMs training two models: a sinus-only VAE (VAE-S) and a class-conditioned VAE (VAE-C) trained on sinus rhythm and AF. Both models were evaluated through morphological, spectral, and distributional similarity metrics. VAE-S achieved higher fidelity with real signals, while VAE-C enabled rhythm-specific generation at the cost of sinus quality. As a proof of concept, we also assessed data augmentation in a downstream task: noninvasive EGM estimation, where moderate inclusion of synthetic signals improved performance. These results demonstrate that VAEs can generate physiologically plausible atrial EGMs, offering a promising tool to alleviate data scarcity in noninvasive EGM estimation.