SAX Mask Adjust

Every workflow that uses a segmentation mask eventually hits the same problem: the mask you have isn't quite the mask the next node wants. Your detailer needs the face region plus a bit of margin; your noise injector needs the face region minus a bit; your finisher needs everything except the face. Running SAM 3 three times with three different grow settings is expensive and fiddly. SAX Mask Adjust is the cheap answer: run the segmentation once, then fan the result out through a couple of these nodes, each tuned to its own downstream job. It's a single-purpose node - grow/shrink, blur, threshold - and it does that one job well.

How it works

One MASK in, one MASK out, with four operations applied in a fixed order: invert → grow → blur → threshold.

• invert - flips "white = target" to "white = protected". The killer feature. Detect the face with SAM 3, invert, and the mask now says "everything except the face" - refine the background while leaving the face alone, no re-segmentation.
• grow (px, −256 to +256) - hard-edged dilation (positive) or erosion (negative). Fast, jaggy, exactly what a hard mask operation should be.
• blur (sigma px) - gaussian softening.
• threshold (0–1) - binarize after the blur. 0 keeps a soft mask.

The combinations are where it gets useful, straight from the pack docs: grow=+8 for a fast hard dilation; blur=3, threshold=0.1 for a smooth, jaggy-free dilation; blur=3, threshold=0.9 for a smooth erosion; and blur=3, threshold=0 for a soft feathered mask - which is exactly what a sampler's noise_mask wants.

The usage pattern the node was built for

The README's example is the canonical one:

SAM3 (mask_grow=0) ──┬─→ Mask Adjust (grow=+8) → Detailer (larger region)
                     └─→ Mask Adjust (grow=-2) → Noise Injector (smaller region)

Set the SAM3 segmenter's own mask_grow to 0 and do all the per-branch tuning here - the two grow values are additive, so mixing them confuses the picture. Fan out with clean intent: one segmentation, one Mask Adjust per branch, each with its own grow/blur/threshold.

The invert use-cases are worth spelling out because they're the thing people don't expect: SAM3 returns background instead of the subject you wanted → flip invert and it's corrected instantly. Want to protect a region and expand protection slightly outward → invert + grow=+8. Face detected but you want everything-but-face detailed → SAM3(face) → Mask Adjust(invert) → Detailer.

What you don't need to worry about

There's no re-running of the model, no prompt, no seed - this is pure pixel math on the mask, instant on any machine. The grow uses the same separable max-pool as the SAM3 segmenter itself, so the semantics match what you already expect from the segmenter's grow field. It's the rare node that's hard to misuse: connect a mask, tweak two or three values, and the output is predictable.

Installing

Part of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/so16tm/SAX_Bridge

SAX/Bridge/Mask → SAX Mask Adjust. Pair it with a SAM3 Loader + Multi Segmenter upstream and a Detailer downstream and you've got the pack's whole text-mask-to-refine loop in four nodes.