In total knee arthroplasty, the femoral implant size is chosen mainly based on the femoral anteroposterior (AP) height and mediolateral (ML) width. This choice often is a compromise, due to limited size availability. Inadequate AP fit is expected to alter flexion laxity and thus knee function. Inadequate ML fit entails underhang or overhang, which is linked to worse clinical outcomes. Hence, we aimed to find implant size distributions, which maximize population coverage, and to evaluate the sensitivity regarding error bounds and the number of implant sizes for a database of 85,143 cases. All patients in the database have been provided with a patient-specific implant in the past. For a subset of 1049 cases, three-dimensional preoperative bone surface models were available. These were used to validate whether the implant dimensions were representative of the bone dimensions. Particle Swarm Optimization was used for optimizing the implant size distribution. The deviations between implant and bone measures in the subset were found to be clinically irrelevant. Therefore, the full database of 85,143 cases was used for further analyses. A higher sensitivity of the population coverage regarding the error bounds compared to the number of implant sizes was found. For an exemplary setup of 12 optimized implant sizes and error bounds of ±1.5 mm for AP and ±3 mm for ML, a population coverage of almost 85% was achieved. In contrast, even with 30 implant sizes, a full population coverage could not be reached. Hence, remaining cases should be provided with patient-specific implants.