RISE speeds up edge-device diffusion models 2.1x without quality loss

Deploying text-to-image diffusion models at the network edge has typically forced a trade-off between large edge-side models offering high fidelity but high latency and lightweight device-side models that sacrifice semantic coherence for speed. RISE (Relay Inference and Online Scheduling) proposed by researchers from several universities breaks this compromise. The key insight is that latent intensity exhibits minimal deviation after a model handoff within the same model family. RISE leverages this with a training-free relay mechanism: the large model on the edge handles the early denoising steps that define semantic structure, then passes the intermediate latent representation to a small device-side model for additional detail refinement. This hybrid approach avoids the need for full on-device inference while maintaining output quality.

To make this practical for real-world services, RISE includes a contextual bandit scheduler that dynamically selects the best relay configuration. It considers prompt complexity, user quality/latency preferences, network bandwidth, and real-time load on edge nodes. The system balances these factors to optimize for each request. Evaluated on two benchmarks, RISE's relay mechanism achieves up to 2.1x speedup compared to running the full large model on the edge, with no degradation in semantic coherence or image quality. The context-aware scheduler effectively handles mixed workloads, adapting to varying conditions. This work is set to be published at IEEE ICWS 2026 and represents a significant step toward practical, efficient deployment of diffusion models on resource-constrained edge devices.