Vision is a very rich sensor with a proven critical role in the control of balance. However, it is widely under- used for robotics postural control. This paper presents and compares two approaches, one model-based and one model-free, to ensure stability of the COMAN compliant humanoid robot standing on a moving platform. The model-based approach uses inverse kinematics, while the model-free one relies on a neural network as mapping between sensors and actuators. The sensory information is composed of proprioceptive cues (gyroscope) and visual cues, used separately or together. We present methods of using vision as sensory input without relying on a particular object or feature of the scene, but only on the optical flow. The performance of both approaches are compared systematically in a realistic robotics simulator, for different movements of the platform and using different sensory cues. We aim to see if vision can replace proprioceptive sensors or be fused with them to improve the performance of the stabilizing controller. While both model-based and model-free approaches successfully stabilize the robot, the model-free approach shows better overall performance. Preliminary results on the real COMAN robot are shown