Hush v3.6.0 — texture reconstruction

Hush — the noise-reduction half of the toolkit. Source & issues: amateurmenace/Hush-OpenNR.

Downloads: macOS .pkg (installer) or .zip; Windows .zip (OpenCL, x64). See the README for install steps.

---

The texture-reconstruction release. A real-footage audit found the denoising
core is already strong — so this release builds the other half: after the
noise is gone, put the image's optical character back. There is now a stage
that ADDS high-frequency energy (nothing before this could), grain that is
matched to the noise that was removed instead of a flat mid-gray dither, a
texture path that no longer re-noises the shadows it fills, a fix for the
blotchy color speckle left deep in shadows, and a confidence matte that hands
downstream nodes a map of where the denoiser actually succeeded.

Everything here is off by default, so existing projects render bit-for-bit
identically until you reach for it. All new math is ported across the CPU
reference and the Metal / CUDA / OpenCL kernels at the usual ~2e-5 parity, and
every idea was CPU-prototyped and gated on real 4K night footage before it
shipped.

Added

• Optical Acutance (Refine). The first stage in the whole pipeline that
  raises high-frequency energy: an edginess-gated high-pass on the cleaned
  luma, hard-clamped to the local 3×3 min/max so edges get their slope back
  with zero ringing by construction — sharpness that reads like a lens,
  not a halo. Gated to real edges by the noise field, so it never amplifies
  noise. Measured +7.9% edge slope on the reference clip; flats untouched.
• Shadow Color Cleanup (Refine, the WEAK-1 fix). The residual mid-frequency
  color speckle in deep shadows sits above the ~23 px reach of the chroma
  bands. A wide luma-guided chroma pass averages it — guided by the clean luma
  so it never crosses a real object edge, and kept to a moderate reach so the
  frame-scale lighting gradient (warm vs cool ambient over dark fabric) stays
  put. −70% shadow chroma speckle on the clip with the real color preserved.
• Clean-Confidence matte (view mode). Exports the per-pixel effective
  sample count — how deep the temporal stage averaged — calibrated into
  RGB + alpha. A downstream node can key shadow-lift, grain and local contrast
  off where the denoiser succeeded (high on static, low on the motion the
  gate had to protect). The complement of the existing noisiness matte.
• Grain Fineness (blue-noise spectrum). High-passes the grain toward the
  eye's contrast-sensitivity peak; at matched RMS it reads sharper and hides
  the plasticky look better than full-band grain.

Changed

• Film Grain is now reconstructed, not dithered. Amplitude follows the
  measured brightness-noise curve (loud in shadows, quiet in highlights —
  where real noise and the plasticky waxing actually live) instead of a
  mid-gray parabola, and it is contrast-masked off edges so it never dithers
  over detail or acutance. (Only affects clips that had grain turned on.)
• Luma Texture now cores the noise out first. The re-injected texture is
  soft-thresholded at the input-noise scale, so it puts back real
  micro-structure without pushing the removed noise back into the shadows —
  measured shadow re-noising dropped from ~4.7× to ~1.2×. (Only affects clips
  that had Luma Texture turned on.)
• Adaptive Strength (effN-steered spatial): the spatial cleaning is now
  spent per pixel where frame-averaging couldn't help (moving subjects) and
  relaxed where it already worked, keeping the residual uniform across motion.
• Auto Setup deep-tune: the self-tuner now also runs a coordinate descent
  over the motion/detail/EQ axes and a chroma-SURE pass, with hold-out
  cross-validation so a tune that only fits the probe's own noise is rejected.

Fixed

• Chroma auto-tune on correlated speckle. Plain MC-SURE is only a valid
  error estimate for white noise; on the spatially-correlated shadow chroma
  speckle it read the win backwards and de-tuned a good profile by ~0.9 dB.
  The probe is now matched to the measured chroma-noise correlation length, so
  the estimator becomes generalized SURE and tracks true error — turning that
  0.9 dB loss into a 3.0 dB gain on the correlated case.