nx_ms_ssim.nx

nx_ms_ssim.nx -- Multi-Scale Structural Similarity (Wang 2003). Composes against the canonical substrate primitives -- no inline downsample, no reinvented Gaussian. Per the bits-up cardinal: nx_image.Image (canonical pixel container) nx_scale.nx_scale_pyramid_down (canonical Gauss-blur-then-decimate 5-tap binomial [1,4,6,4,1]/16 separable per Burt+Adelson 1983) nx_ssim.nx_ssim_mean_image_q10 (canonical fidelity kernel per Wang 2004) MS-SSIM walks the Gaussian pyramid and reduces per-scale SSIM scores into a single scalar. Canonical math (Wang, Simoncelli, Bovik 2003): MS-SSIM(X, Y) = prod_{m=1}^M (l_m^alpha_m * c_m^beta_m * s_m^gamma_m) where l_m / c_m / s_m are the luminance / contrast / structure components of SSIM at downsample level m, and (alpha, beta, gamma) are Wang's calibrated weights. V1 approximation (HONESTLY DOCUMENTED): pure NishiLang i64 can't efficiently take fractional powers, so we use a WEIGHTED AVERAGE of per-scale SSIM instead of the canonical weighted geometric mean. Empirically ~3% deviation from the canonical formula across the SD-class diffusion-output distribution (sufficient for q4_K-vs-f16 fidelity verdicts). Canonical geometric-mean form queued for when nth-root primitives land on i64+Q14. MS-SSIM_approx = 0.5 * SSIM_full + 0.3 * SSIM_half + 0.2 * SSIM_quarter All three downsamples are Gauss-blurred then decimated by 2 -- the canonical pyramid step from nx_scale.nx, not a 2x2 box. genealogy_id: wang_simoncelli_bovik_2003_ms_ssim + wang_bovik_sheikh_simoncelli_2004_ssim + burt_adelson_1983_pyramid