Human-to-robot handovers play an important role in collaborative tasks in industry or household assistance. Due to the vast amount of possible unknown objects, learning-based approaches gained interest for robust and general grasp synthesis. However, obtaining real training data for such methods requires expensive human demonstrations. Simulated data, on the other hand, is easy to generate and can be randomized to cover the distribution of real-world data. The first contribution of this work is a dataset for human grasps generated in simulation. For this, we use a simulated hand and models of 10 objects from the YCB dataset [1]. It can also be easily extended to include new objects. The method thus allows for generating an arbitrary amount of training data without human interactions. Secondly, we combine a generative neural grasp generator with an evaluator model for grasp pose generation. In contrast to previous works, we obtain grasp poses from simulated RGB images which allows for reducing the negative effects of depth sensor noise. To this end, our generator model is provided with a cropped image of the human hand and learns the distribution of grasps in the wrist system. The evaluator then narrows down the list of grasps to the most promising ones. The presented approach requires the model to extract relevant features from images instead of point clouds. A cost-efficient method for generating large amounts of training data is therefore needed. We test our approach in simulation and transfer it to a real robot system. We use the same objects as in the training dataset but also test the generalization capabilities toward new objects. The presented dataset is available for download: https://tuc.cloud/index.php/s/g3noZD7oCqbQR9d .