Lightweight deep learning rendering framework for edge-deployed virtual reality art therapy systems

Real-time neural rendering of emotionally expressive visual content is central to virtual reality art therapy, yet deploying such models on resource-constrained edge devices remains an unsolved engineering challenge because of strict latency, power, and memory limits. This paper proposes a lightweight deep learning rendering framework tailored to edge-deployed VR art therapy systems, positioned as a system-level co-design rather than a new algorithmic component. Three interdependent modules make up the framework. First, a hybrid model compression strategy couples channel pruning with knowledge distillation under a single joint loss and, as an added ingredient, weights the pruning importance score by a therapy-content saliency map so that channels supporting therapeutically salient regions are protected. Second, a therapy-content-aware adaptive rendering quality controller allocates computational resources through semantic saliency prediction (no runtime gaze tracker is used in the current prototype). Third, a three-stage pipelined inference architecture overlaps preprocessing, the neural forward pass, and post-processing across heterogeneous edge processing units, with a feedback loop from pipeline monitors back to the saliency threshold. Experiments on three edge platforms and five art therapy scenes of varying complexity show that the framework achieves a throughput latency of 7.1 ms per frame on the Jetson Orin NX, comfortably meeting the 11.1 ms per-frame budget for 90 Hz rendering, while retaining 94.3% of the uncompressed teacher’s structural similarity. Objective perceptual metrics (VMAF, FLIP, DISTS) and subjective mean opinion scores collected under a double-blinded protocol from thirty raters (fifteen computer graphics researchers and fifteen with clinical or art-therapy training) confirm that colour fidelity and temporal coherence are preserved. The results concern engineering feasibility rather than clinical efficacy; therapeutic benefit is hypothesised, not demonstrated, and would require patient-level trials that lie outside the present scope.