ControlMask: per-pixel control of NVIDIA NR

Back to the report · Static analysis · Noisy-domain NR feasibility

Written by Claude Opus 5.5 (Claude Code, high effort) on behalf of GrimsVerk, who skimmed it. The research, the test-build code changes, the measurements and this text are Claude's work.

What the runtime does with it

DLSSNR.ControlMask (+ ControlMaskSubrectBase/Width/Height) is read at every evaluation, in both 310.8 and 310.9.16. Per pixel:

Channel	Effect
r	intensity: the final blend becomes out = orig + sat(Intensity × mask.r) · (NR − orig)
g	tone: network input channel 11 = LocalTone × mask.g. The Style colour grade is not scaled by it (the grade uses clamp(LocalTone, 0, 1))
b	structure: AutoMask off → channel 12 = structure × mask.b; AutoMask on → see MaskBlendMode
a	unused

Interaction with NVIDIA's automatic skin/structure mask (AutoMask):

310.8: any ControlMask forces AutoMask off.
310.9.16: DLSSNR.MaskBlendMode.
- 0: as in 310.8, AutoMask off.
- 1 / 2 / 3: AutoMask stays on, and m = mask.b × structure is combined with NVIDIA's skin (channel 13) and structure (channel 14) values by average / min / max.

With a mask present, the runtime uses the ..._control_mask variants of the fused final kernel (..._full_rect when the subrect covers the whole texture). See the static analysis, section 4.

How the research build makes the mask

The research build is based on GrimsVerk's OptiScaler-NR fork v0.3.8, commit ac313d06; the source is in research-build-source.patch.

Where: a new mode (14) of the fork's NR compute shader writes an RGBA8 mask at the model's working size every frame, right before NVIDIA's evaluate. The mask is passed with a full-texture subrect; null when the mask is off.
Inputs:
- t0 = the picture the model gets;
- t2 = the game's depth (guide resolution);
- t3 = the motion vectors.
Rules, applied in order on top of a base value:
1. Far / sky: a soft weight, full at depth ≤ margin and zero at 2 × margin (reversed depth; mirrored for ordinary depth). It is averaged over 5×5 depth texels so silhouettes do not flicker with the per-frame jitter. Blends the base towards the far values.
2. Dark: below "dark start" brightness (fully at "dark full"), r and b fade towards the dark values.
3. Motion: between "motion start" and "motion full" pixels per frame (at the model's size), r and b fade towards the motion values.
Cost: +0.13 ms at 4K output with the small model (mask build + NVIDIA's mask kernel variant, measured inside the model timer: 7.26 vs 7.13 ms).

Settings (OptiScaler.ini, section [DlssNr])

Key	Values	Meaning
`ControlMask`	true / false	build and send the mask (D3D12 path; off in spatial-compression mode)
`MaskBlendMode`	0-3	see above. 2 recommended: NVIDIA's skin handling stays; the mask can only lower structure
`ControlMaskParams`	16 numbers	see below

ControlMaskParams order, in four groups of four:

Group	The four numbers
1	base r, base g, base b, (unused; the depth direction is filled in automatically)
2	far r, far g, far b, far margin (0 = off; 0.00001 = sky only in reversed-depth games; 0.001 = also distant scenery)
3	dark start, dark full (brightness 0-1, start > full), dark r, dark b
4	motion start, motion full (pixels per frame at the model's size, full > start), motion r, motion b

The default 1,1,1,0, 1,1,1,0, 0,0,1,1, 0,0,1,1 changes nothing. The same values are available as live test-control keys (ControlMask, MaskBlendMode, MaskParam0…MaskParam15) when HudlessTestControl = true. When the mask is created, OptiScaler.log shows a line DLSS-NR control mask WxH: base …, far … margin …, dark …, motion …, MaskBlendMode ….

Example, keep the game's own sky:

ControlMask = true
MaskBlendMode = 2
ControlMaskParams = 1,1,1,0, 0,1,1,0.00001, 0,0,1,1, 0,0,1,1

Example, less NR and structure in the darks: ControlMaskParams = 1,1,1,0, 1,1,1,0, 0.20,0.04,0.4,0.5, 0,0,1,1

Proof (The Witcher 3, frozen photo-mode frames)

Each value is the mean |difference|, in 8-bit units, between the two captures named in the first column. The "noise floor" row is the same setting captured twice.

Comparison	Courtyard 310.9.16	Yennefer 310.9.16	Courtyard 310.8	Yennefer 310.8	Meaning
def_end vs def	2.03	1.41	1.55	1.37	noise floor (same setting twice)
m1_b0 vs def	2.04	1.49	1.62	1.50	all-ones mask, MaskBlendMode 0: no visible change
m1_b2 vs def	2.03	1.41	-	-	all-ones mask, MaskBlendMode 2 (min): no change
mr05 vs int05	1.57	1.36	1.47	1.41	red = 0.5 is the same as Intensity 0.5
mr0 vs def	7.63	7.61	-	-	red = 0: NR fully off where the mask says so
mb0_b0 vs str0	1.44	1.42	-	-	blue = 0 (blend 0) is the same as Local structure 0
mb0_b1 vs def	3.04	2.13	-	-	blue = 0, blend 1 (average with NVIDIA's skin/structure mask): half the effect
mb0_b3 vs def	2.03	1.41	-	-	blue = 0, blend 3 (max): no change, NVIDIA's own value wins
mg0 vs def	4.89	3.48	5.84	4.65	green = 0: tone input off; the style grade stays (unlike Local tone 0)
mg05 vs tone05	2.37	2.87	-	-	green = 0.5 against Local tone 0.5: close, not identical (grade)
mdark vs def	2.50	2.17	-	-	dark rule (less intensity/structure in the darks)
tone05 vs def	3.65	3.25	-	-	for reference: Local tone 0.5 (no mask) against 1

Sky rule (terrace frame with a large sky):

Edit maps: NR's edit in the sky area drops to nothing; a margin of 0.001 also removes it from everything beyond the terrace wall.
Flicker: the first hard-edged version raised frame-to-frame flicker from 0.22 to 0.28. The soft 5×5 version brings it to 0.212 against 0.204 without a mask.
WorkingScale 0.5 (from the INI only): the edit in the top third of the frame dropped from 12.3 to 4.8; the rest is mountains and a tree.

Pictures are in the ControlMask section of the report.

Caveats

The motion rule is untested, because photo mode has no motion.
The mask is 8-bit, and it multiplies the global Intensity, Tone and Structure settings rather than replacing them.
310.8 has no MaskBlendMode, so any mask there switches AutoMask off; for an all-ones mask, the measured effect is at the noise floor.