MMEdge: Real-time Multimodal Inference via Cross-Modal Speculative Sensing and Encoding
Runxi Huang, Xiaomin Ouyang · 2025
Achieving real-time multimodal inference on devices is crucial for time-critical and privacy-sensitive applications. Existing methods primarily focus on optimizing the latency of model inference while neglecting the data collection process. We introduce MMEdge, an on-device, end-to-end multimodal inference framework. MMEdge employs a cross-modality speculative sensing and encoding approach to enhance inference efficiency and robustness at runtime.