Latency-efficient cloud-edge cultural pattern generation via lossless lightweight condition transmission and bandwidth-adaptive scheduling

Abstract Cloud-edge systems for controllable diffusion generation typically upload the raw reference image, although ControlNet consumes only a sparse structural condition derived from it. We present a cloud-edge framework for directed cultural pattern generation in which the edge extracts the condition map, compresses it losslessly, and uploads only this compact representation, while the cloud runs the full Stable Diffusion and ControlNet pipeline. Joint latency, energy, and quality models yield a bandwidth-adaptive policy over condition type and compression level. On a curated set of 112 cultural-pattern images (81/31 train/held-out split, 1–10 Mbps uplinks), condition transmission reduces the mean uplink payload from 227.0 KB to 14.6 KB (about 94%) and the mean transmission-pipeline overhead from 0.8557 s to 0.0865 s (about 90%) relative to raw-image upload, while the generated images remain pixel-identical to those produced from the uncompressed reference, verified end-to-end (identical FID, LPIPS, and CLIP scores). PNG compression level itself proves to be a useful scheduling variable: fast PNG-L3 overtakes maximum-compression PNG-L9 above 2 Mbps and balanced PNG-L6 above about 8 Mbps, so no single fixed configuration is optimal across heterogeneous uplinks. A hybrid-action reinforcement-learning scheduler trained on real per-sample measurements selects the transmission configuration automatically: it matches the strongest fixed-branch and discrete-action baselines on every reported metric, crosses the measured codec crossover without any hand-set threshold, and, unlike discrete schedulers, can absorb continuous codec controls without re-discretization. Reported latencies are transmission-pipeline overhead excluding cloud diffusion; an end-to-end budget with measured generation time is also provided.