Introduction. Atrial fibrillation (AF) is a prevalent arrhythmia for which current treatments remain suboptimal. Digital Twins (DTs) enable patient-specific modeling to guide therapy, however many existing frameworks rely on invasive data or recordings outside AF episodes. This work introduced a non-invasive framework to generate DTs using electrocardiographic imaging (ECGI) recorded during AF. Methods. DTs were personalized using dominant frequency (DF) maps obtained from ECGI recordings acquired during AF. These maps guided the adjustment of action potential duration (APD) restitution curves at each point in the atria through an optimization process. The curve shape was adapted based on DF and its temporal variability, assigning flatter restitution curves to regions with high DF and low variability—indicative of more stable and leading behavior—and steeper curves to less regular areas. A grid search over restitution and conduction parameters was performed, and the simulation with the lowest DF error was selected as the final DT. Results. In five patients, the framework capacity to capture the DF gradients leads to an R² =0.97 when comparing the 5th and 95th percentiles of DF between the patient and the DT. Moreover, the selected DTs showed a mean absolute error of 0.21 ± 0.03 Hz and a correlation coefficient of 0.77 ± 0.08 compared to patient DF maps. In all cases, DTs reproduced the clinical ablation outcomes, correctly reflecting AF termination or persistence. Conclusions. This framework generates Digital Twins that replicate patient-specific AF dynamics using only non-invasive data acquired during the arrhythmia. Its real-time applicability makes it suitable for integration into clinical workflows to support personalized ablation strategies.