← Features
Masking

Guided-Filter Feathering

Soft mattes that respect edges. Transduce feathers with a guided filter — the same edge-preserving family used in other leading software packages — so softness follows real detail instead of a blind Gaussian haze.

Most grading tools still soften a qualifier with a plain blur: every pixel gets the same falloff, whether or not there is hair, fabric weave, or a hard contrast edge underneath. That is fine for a soft vignette. It is wrong for a skin matte that has to stop at frizz, or a sky key that has to cling to a boom-mic silhouette.

Transduce replaces that last-stage blur with a guided filter driven by the frame’s own luminance. The algorithm softens where the image is flat, and tracks structure where the guide has real variance — standard technique, production-grade behaviour.

Why it matters

Gaussian haze vs edge-aware softness

Blind blur spreads radially. Guided feathering bends around the plate.

Side-by-side concept: Gaussian matte haze vs guided-filter feathering that clings to luminance edges
Concept — left: uniform Gaussian smear. Right: guide-aware falloff that preserves fine boundary detail.
Industry standard algorithm

The same technique the biggest tools lean on

Guided image filtering (He, Sun & Tang — ECCV 2010 / TPAMI) is not a Transduce invention. It is a published edge-preserving filter with a linear-time path — the same class of technique used in other leading software packages for guided feathering and edge-aware matte refine.

How it shows up
Other leading software packages Edge-aware / guided feather controls that soften a matte while following structure in a guidance image
Transduce Built into the qualifier refine stack as the default feather — guide = frame luminance, no extra node graph required

We cite the technique honestly: matching what colourists and compositors already trust elsewhere — not claiming we own the paper.

How Transduce applies it

Inside the Mask panel’s refine column — after the selection is made, before intensity and invert.

Despeckle→ Hole Fill→ Expand / Contract→ Guided Feather→ Intensity→ Invert
Detail Behaviour
Guide image Per-frame luminance from the source — already available in the qualifier path, no second render pass
What softens The matte only — colour grade stays on whatever the selection encloses
Edge behaviour High local variance in the guide preserves structure (hair, fabric, hard lines); flat regions blur smoothly
Implementation Standard He et al. linear model from box filters — no proprietary ML, no optional image-processing addon required
Fallback If a guide frame is not available to a caller, feather falls back to Gaussian so the control never silently no-ops
Order reason Expand/contract runs before feather — closes noise-driven micro-holes while the matte is still hard-edged; guided feather then softens one clean boundary instead of fogging hundreds of speckles

Why analytical feathering still wins on AI plates

Qualifier + guided feather is computed fresh from pixels every frame — no learned segmentation, no matte flicker from a model that drifts shot to shot. That temporal stability is the point of the masking engine on DNA-06; this page is the soft-edge half of that story.

Note: geometric shape soft-edge uses a distance ramp (straight falloff on the shape boundary), not the guided filter. Guided feathering applies to the qualifier matte refine path.

Related

Edge-awareGuide = frame luminance.
Published methodHe, Sun & Tang guided filter.
Refine-readyDespeckle → fill → grow → feather.
StableAnalytic mattes, no ML flicker.

Soft edges that know where the plate ends.

Industry-standard guided filtering — wired into the qualifier so hairlines stay hairlines.

Download for macOS (Apple Silicon) Available for macOS on Apple Silicon only. Windows & Linux coming soon.