Deployment of video communication services over mobile ad hoc networks faces numerical obstacles since the networks are characterized by limited and unpredictable bandwidth. Following the tenet of service-orientation, this study proposes a cross-layer design for fine-grained adaptive transmission of real-time video over time-varying channels. Specifically, the sender accurately estimates the current effective bandwidth by observing its transmission buffer and negative acknowledgments (NACKs) from the receiver. Information on route quality provided by routing protocols may also be exploited to control the real-time stream efficiently. At the same time, the sender handles NACK messages to predict route goodness and orders retransmissions for improving video quality. We have evaluated the performance of the proposed design by both simulations and experiments of ad hoc nodes delivering H264/AVC, which consistently demonstrated the soundness and feasibility of the framework with respect to video quality and resource utilization.