NV3D Vision Restoration Project

Timing, jitter and illumination calibration

Written 2026-09-13. This describes how the app times the shutters against the display, what it measures about that timing, and how to align the glasses with a panel that strobes (LightBoost-style backlight, black frame insertion, “Motion Clearness”). The numbers below were measured on this PC with the Hisense U6 Pro at 4K 143.988 Hz; they are host timings, not optical measurements.

The timing chain

Presentation timestamps. For every refresh DXGI reports the refresh counter and a QPC timestamp (DXGI_FRAME_STATISTICS::SyncQPCTime). The presenter fits a straight line through the last 240 of them (least squares, TimingTracker). The line gives the period and the vblank phase; each new timestamp’s distance from the line is the vblank jitter shown under Diagnostics. Earlier builds anchored the prediction on the newest timestamp alone, so its whole jitter went into the next command.
Predicted vblank. The eye command for the refresh a present will land on is due at the fitted time of that refresh. The USB worker waits on a high-resolution timer and sends the 8-byte eye command; the difference between the moment the write starts and the deadline is the eye command timing error (last / rms / max under Diagnostics).
Emitter. The command carries the distance to the emitter’s next period boundary; the X register delays the open token after the boundary and Y closes it. The phase control is split into those two values (nvidiaSchedule), every phase or shutter change rewrites the timing block (counted as timing writes), and the Output & timing tab shows the boundary and X the emitter currently receives.
Panel. Rows are rewritten top to bottom over the scan time, settle over the panel response, and are lit continuously (sample and hold) or only during a pulse (strobed).

Measured on 2026-09-13 (Hisense, 4K 143.988 Hz)

Quantity	Value
Presentation timestamp jitter around the fitted line	0.9 to 6 us rms, 4 to 24 us max
Scan-line estimate of the vblank start (kernel counter after a vblank wake)	22 to 27 us rms
Wake latency of the vblank event	140 to 150 us median, up to 400 us
Blanking interval (154 of 2314 lines)	462 us
Presentation timestamp after the blanking interval starts	105 to 150 us (varies by run)
Scan of row 0 after the presentation timestamp	105 to 353 us (varies by run)
Mean USB transfer of an eye command	about 120 us

The DXGI timestamps are therefore the better clock, and the scan-line probe is used only to calibrate where the scan starts relative to that timestamp and to report the blanking length. Diagnostics > Measure vblank (3 s) or VisionRestoration.exe --vblank (reports/vblank.txt, runs beside an open app) performs that measurement and stores the scan start offset in the profile (Advanced > Panel). The standalone probe that produced these numbers is in experiments/vblank-probe.

Frame jitter under Diagnostics

Vblank jitter rms / max over the fitted refreshes: display timestamps against the refresh clock. Values above about 50 us rms mean the clock is not trustworthy (a composed window, a mode change, power saving).
Present interval jitter: standard deviation of the presenter’s present-to-present interval over the last 256 frames. Occasional two-period intervals show up as slips in the presentation counters, not here.
Eye command timing error: how far from the predicted vblank the USB write started. This is host scheduling; the emitter’s own period lock filters it further.
Mean USB transfer is subtracted when a phase is suggested from the model.
All of these are written to reports/session.log every two seconds and to the exported timing report.

The illumination model

estimateLeakage follows 33 rows of the stereo area through one stereo cycle. Each row receives a refresh’s content when the scan reaches it, fades to it linearly over the panel response, and is lit continuously or only during the strobe pulse. A shutter that is open from the phase for the shutter length collects the eye’s own light and the other eye’s light; their ratio is the predicted leakage, reported for the top, center and bottom third of the area. The Stereo area section shows this prediction for the current phase; Suggest phase searches the whole cycle for the least leakage (brightest among equals), and Fit area to shutter picks the largest band whose best phase leaks at most 1 %. Black frames and repeated frames go through the same model.

Model phases are measured from the start of the eye’s first scan. The emitter phase is measured from the presentation timestamp plus the USB latency, and for Left/Left/Right/Right from the second refresh. emitterPhaseFromModel converts between the two, using the measured scan start offset.

What the model says for the displays on hand:

Configuration	Best whole-screen result
4K 120 Hz Left/Right, OLED (scan 8.1 ms)	no clean phase; about 7 % leakage in the top and bottom thirds with a 1.8 ms shutter, center clean
4K 120 Hz Left/Right, Hisense (scan 6 ms, response 3 ms)	worse; use the stereo area
4K 240 Hz Left/Left/Right/Right, OLED	clean window of about 4 ms
4K 240 Hz Left/Black/Right/Black, OLED	clean for shutters up to about 5 ms, 84 % of a fully lit frame once the shutter is widened, 60 flashes per eye per second
Strobed panel, pulse after the rows settled	clean whenever the shutter covers the pulse

Software black frame insertion

Output & timing > Software black frame insertion (or B in fullscreen) switches the sequence to Left / Black / Right / Black. While a black frame is scanned in, no row anywhere on the screen shows the other eye, so the whole screen can be clean on a sample-and-hold panel. The costs are a per-eye flash rate of a quarter of the refresh rate (60 per second at 240 Hz, which is what 3D Vision delivered at 120 Hz, but only 30 per second at 120 Hz, which flickers) and less light, most of which the wider shutter described under “Getting the light back” recovers. It is a test tool at 120 and 144 Hz and a real route at 240 Hz.

The game hook must present four frames per game frame for this. The shared-memory contract carries the sequence (sequence field) and a hook advertises support with the HookSequences flag; the current hook build does not, so the app drives the game as Left / Right whatever the profile says and reports that in the hook status.

Getting the light back

A black-frame or repeated sequence lights each eye on fewer refreshes, so it is dimmer. How much of that is recoverable is worth stating precisely, because two different effects look the same through the glasses:

Inherent. A row is lit with its own eye for one refresh out of the four-refresh cycle. No shutter setting changes that.
Recoverable, and previously thrown away. The emitter runs one period per stereo slot pair, so a four-slot sequence gives it a period of two display refreshes, and the shutter may stay open for that whole period. Every shutter limit in the app used to be written against one display refresh, which cut the usable window roughly in half (and at 240 Hz to a sixth). A row shows its eye from the moment the scan reaches it until the next scan one refresh later, and the other eye cannot appear before two refreshes, so a shutter of about period + scan, opened at the scan start, collects every row’s entire lit period and still never sees the other eye.

Maximize brightness (next to Suggest phase) does this: it widens the shutter to the largest value whose best phase stays within 1 % leakage and recentres the phase. It ranks candidates by the light actually collected, not by the lit fraction of the shutter, which falls as the shutter widens and would have chosen the dimmer setting. The stereo area reports the result as “light reaching this eye” against one fully lit frame.

Configuration (whole screen, model)	Old cap	New cap	Light before	Light after	Leakage
OLED 4K 120 Hz Left/Black/Right/Black	4980 us	13276 us	51 %	96 %	0 %
OLED 4K 240 Hz Left/Black/Right/Black	813 us	4980 us	19 %	84 %	0 %
OLED 4K 240 Hz Left/Left/Right/Right	813 us	4980 us	19 %	118 %	0.6 %
Hisense 144 Hz Left/Black/Right/Black	3592 us	10499 us	46 %	95 %	0 %
Hisense 120 Hz Left/Black/Right/Black	4980 us	13276 us	55 %	98 %	0 %

Above 100 % is real for Left/Left/Right/Right, where a row is lit for two refreshes. Pressing Maximize brightness on a whole-screen Left/Right configuration does the opposite and shrinks the shutter hard, because that is the only way that sequence stays inside the leakage limit; the notice reports what it chose.

What is left after the shutter is as wide as it can go is panel luminance. Image brightness (Output & timing) multiplies the image the app presents. In HDR it reaches into the display’s headroom; in SDR it clips to white, so use the display’s own brightness control instead. It applies to the app’s own output only, never through the game hook, and it can never lift a black frame.

Strobed panels (LightBoost, BFI, Motion Clearness)

A strobed panel lights the whole screen at once for a short pulse after the scan. If the rows have settled by then, the clean window is the pulse itself and it covers the whole screen: this is how the original 3D Vision monitors worked. The pulse timing of a TV’s BFI mode is not published, so the app finds it with the glasses:

Enable the display’s BFI (Hisense: Picture > Motion > Motion Clearness; Samsung: Clear Motion). Some TVs offer it only at 60 or 120 Hz input.
Advanced > Panel: set Illumination to strobed. Set the shutter to its minimum (250 us) so the glasses act as a narrow probe.
Start fullscreen 3D with Both eye targets and press S. The phase now advances through the whole cycle over 30 seconds (speed selectable under Output & timing) and its value is drawn in the output. Through one lens the targets are dark most of the time and bright twice per cycle: once with the correct image (the target on the correct side, the correct LEFT/RIGHT label) and once with the other eye’s image.
Press M where the correct image first brightens and M again where it fades. Press S to stop. Set strobe from the last two marks stores the pulse (start = first mark + shutter, end = last mark).
Open the shutter to the length you want (it must cover the pulse; brightness comes from the pulse, not from a longer shutter) and press Suggest phase. The prediction under the stereo area should read 0 % top to bottom. Sweep from there if the lenses’ own delay moves it.

If the bottom of the picture still shows the other eye at the pulse, the panel has not settled when it strobes; the model then reports bottom leakage and a shorter area is the only remedy. A high-speed camera clip of the bare panel (see the 2026-09-13 capture in reports/capture) gives the same pulse timing directly.

Which timing controls act on what

Control	Acts on	Available when
Phase	boundary distance and X in every command; parity of the commanded eye	any connected emitter
Shutter	Y register (per-frame on NVIDIA); limited by the emitter period, which is two display refreshes in a four-slot sequence	any connected emitter
Image brightness	the app’s own presented image, before the readouts, never a black frame	always; effective in HDR, clips in SDR
Maximize brightness	the shutter and the phase together	panel scan known
Left / right shutter, right eye offset	per-frame X/Y rewrite	NVIDIA emitter, Left / Right sequence only
Sequence, software black frame insertion	presenter cadence, emitter period	NVIDIA emitter (RP2040 is Left / Right only)
Stereo area	presenter and hook shaders	always
Panel response, scan, illumination, strobe, scan start	the model only (Suggest phase, Fit area, predicted leakage)	always
Phase sweep, marks	the phase, ten steps per second	3D output running
Match output rate, manual rate	emitter period and scaled timing	always

Controls that cannot act in the current configuration are disabled and say why. A simulated emitter reports that its timing controls change nothing physical.

This site is open source. Improve this page.