HERMES-SR: A Lookup-Augmented Recurrent Operator for Live-Camera Super-Resolution

Two constraints define mobile super-resolution and neither matters on desktop. The first is the operator-coverage envelope of the neural-engine compiler, which silently drops Fast Fourier transforms, deformable convolution, per-pixel matrix inverses, and most of the primitives that contemporary image-restoration papers take for granted. The second is on-chip cache geometry: roughly a megabyte of auxiliary memory before LPDDR5X round-trips start to dominate latency. Mobile SR has answered each of these with a separate architectural family — anchor-plain pixel-shuffle networks, reparameterized edge-oriented blocks, cascaded lookup tables, recurrent feature warping — but no published design has combined all four inside a fifteen-frames-per-second budget. HERMES-SR is that combination: roughly 120K parameters, 8 GMAC per output frame, INT8-clean throughout. On the Samsung Exynos 2400 with its Xclipse 940 GPU and revamped neural engine, the projected operating point is 32.0 dB PSNR at 1080p→4K (×2) and 30.0 dB / 0.87 SSIM at 360p→1080p (×3, joint denoise), each at roughly 55 ms per frame and 600 mW of sustained draw. All performance numbers in this preprint are projections derived from the published mobile super-resolution literature. They are targets for future device measurement, not measured results.