perseval-BLR/NeuralScreen

So you want me to send you the log file with the different settings? If you use discord you can tell me what you need so i can do it more efficiently

My discord is the same as my user here

Thanks for the report - NVOFA and CPU DIS are two different estimators, and the difference you see is real, not your setup.

**Why they do not behave the same**

DIS runs inside the program, and its vectors pass two tests before they reach the neural network:

- a **noise floor** - vectors shorter than 0.5 px are zeroed, because DIS reports small vectors even on a still screen (capture noise, cursor jitter), and the network would smear text and UI on them;
- a **static hypothesis** - a vector of any length is dropped unless warping the previous frame by it explains the pixel better than standing still does. This is the test that catches a wrong vector on static text next to a moving edge.

NVOFA runs in the display driver, on the worker side. When it reports active, the program stops computing motion entirely and hands the driver's field straight through - no noise floor, no static hypothesis. Those two tests exist only on the DIS path. So the driver's raw output, including its errors,...

tested on **v1.13.1**

**[Completed]**
- 'Save as' and auto save work nicely

**[Bugs]**
- Switching from HDR display to SDR resulted in endless load screen (HDR option was still on).
- The App popup/menu (in settings) is flickering (whole window, normal - full black window) after pressing Select screenshot folder, and moving the cursor while the choose folder dialog is on. Happened on SDR display with hdr off option, no such bug on HDR display with hdr option on.
- Flickering persists, i have attached a video aswell
- Auto menu off on 'NR On' is nice but if possible to have this as an option (sometimes you want to customize the effect). Also, on the TV (with settings hdr on), after NR On > popup dissapeared > Menu toggle (num2) moved cursor to a position at center up on screen (no menu) > toggled several times with the same bug. The menu appeared as it should once i turned off the NR (num1).
- Some screenshots resulted in black screen with just the app menu in view (attached bellow), ...

v1.13.0 is out: [NeuralScreen v1.13.0](https://github.com/perseval-BLR/NeuralScreen/releases/tag/v1.13.0).

Two things from your report are fixed there:
- **Save As in the screenshot:** the processed frame is frozen before the dialog opens, so the dialog cannot appear in the saved image.
- **SDR flicker on the HDR TV:** SDR capture now goes through the original `DuplicateOutput` converted to a stable BGRA8 (your log showed 1,991 FP16/BGRA8 switches); `DuplicateOutput1` is reserved for HDR. Local smoke on the SDR path: 12 frames, no format-bridge rebuilds.

Please retest on **v1.13.0** and send a fresh log if the flicker or the brightness anomaly remains - HDR brightness across GPUs is the part that is still yours to confirm, not something I can reproduce locally.

v1.13.0 is out: [NeuralScreen v1.13.0](https://github.com/perseval-BLR/NeuralScreen/releases/tag/v1.13.0).

Your log drove the capture fix: the BGRA-only `DuplicateOutput1` request was not enough (the driver still alternated FP16/BGRA8 1,196 times), so v1.13 converts SDR through the original `IDXGIOutput1::DuplicateOutput` to a stable BGRA8 and keeps `DuplicateOutput1` for the opt-in HDR path only. Local DDA smoke on the SDR path: 12 frames, zero format-bridge rebuilds.

Please retest on **v1.13.0** and check the log for the format-bridge line:
`[cap] capture format ... -> ... - rebuilding the bridge` should not repeat.

If the flicker is gone, I will close this; if it is not, the fresh `NeuralScreen.log` plus a diagnostic package (Program tab) is what tells a capture-format problem apart from a display/compositor one.

v1.13.0 is out: [NeuralScreen v1.13.0](https://github.com/perseval-BLR/NeuralScreen/releases/tag/v1.13.0).

It changes the diagnostic side of the CreateFeature failure rather than claiming to fix it. Before any desktop capture, v1.13 runs an isolated preflight on three synthetic 640x360 frames. The native worker now reports the result of `CreateFeature` explicitly (`CACK`), and a failure, timeout, crash, TDR or device loss is a named, staged outcome with a timed quarantine - instead of a restart loop. Plain `SAFE PASSTHROUGH` pixels no longer count as evidence of support.

If you can, please retest on **v1.13.0** and send:
1. `NeuralScreen.log` from the run (v1.13 writes the preflight stage and the reason);
2. a **diagnostic package** - *Create diagnostic package* on the Program tab. It contains the adapter, driver, stage, HRESULT/SEH/DRED markers and a scrubbed log tail, with usernames and paths removed.

That package is exactly what is needed to tell a runtime refusal apart from a dr...

### Studio Driver (v596.36)
```
[env] NeuralScreen 1.12.0 | Windows 10.0 (build 26200) | Windows-11-10.0.26200-SP0
[env] GPU: RTX 4060 (40xx, arch 0x190)
[env] driver: 32.0.15.9636
[env] HDR: off
[env] Num Lock: on | lang: en | profile: Natural | work_scale: 0.65 | flow: fast
[main] monitor '\\\\.\\DISPLAY1' from config.json is not connected - using monitor 0
[main] monitor 0 is 1920x1080 (config: 3840x2160), taking the real resolution
[main] NeuralScreen - profile 'Natural', resolution 1920x1080, monitor 0
[main] NGX parameters: {'style': 1, 'auto_mask': 1, 'intensity': 1.0, 'local_tone': 0.5, 'local_structure': 1.0, 'skin_structure': -1.0}
[main] work_scale 0.65 (NGX resolution 1248x702)
[main] GPU: RTX 4060 · Ada (group 0x190, officially supported: no)
[main] pre-Blackwell GPU: warmup 120 -> 4 to avoid a false frame-0 watchdog timeout
18:42:45.769 [host] adapter 0: NVIDIA GeForce RTX 4060 vendor=0x10DE vram=7956MB luid=00000000:01820EA0
18:42:45.769 [host] adapter 1: AMD Radeon(TM...

v1.13 does not claim to fix the NVIDIA runtime CreateFeature access violation. It adds an isolated three-frame 640×360 preflight before capture/presentation, fail-closed quarantine and a privacy-scrubbed diagnostic bundle. Please retest v1.13 on the RTX 4060 and attach the diagnostic bundle.

The new log confirms the same v1.12 behavior: 1,991 FP16/BGRA8 switches. v1.13 applies the stronger SDR DuplicateOutput/BGRA8 path; local real-GPU DDA smoke had no format rebuild. The screenshot fix is already confirmed, and v1.13 also adds persistent output folders and separate processing/output/skipped-frame metrics. Keeping this open for an HDR-off TV retest.

Your v1.12 log showed that the BGRA-only DuplicateOutput1 request was insufficient: the driver still alternated FP16/BGRA8 1,196 times. v1.13 now uses the original DuplicateOutput BGRA8 conversion for SDR and reserves DuplicateOutput1 for HDR. A real-GPU DDA smoke completed 12 frames with zero bridge rebuilds. Recording remains fixed; keeping this open until the 10-bit flicker is retested on v1.13.

Thank you for the quick fix and the clarification

- Fixed : Screenshot has the correct frozen frame
- Not Fixed : HDR-off TV flicker. From my testing, it seems that the flicker is not based of NR being on, but the app itself. I turned NR off and the flicker persisted until i close the app. The flicker appears upon switching the display (monitor) in settings to the HDR display (monitor), and disappears if HDR option is turned on or if the program is closed, i tested with other options (Boost and 4.5 FG without any effect).

Here is the log, in case it is of any help

[NeuralScreen.log](https://github.com/user-attachments/files/32263948/NeuralScreen.log)

I checked the latest version; the video now saves correctly, but unfortunately, the flickering hasn't stopped.

[NeuralScreen.log](https://github.com/user-attachments/files/32263706/NeuralScreen.log)

Thanks for the detailed feedback. v1.12.0 is now available: https://github.com/perseval-BLR/NeuralScreen/releases/tag/v1.12.0

Fixed in this release:

- The processed frame is now frozen **before** the screenshot **Save As** dialog opens, so the dialog should no longer appear in the saved image.
- The HDR-off TV flicker in your log was the same DXGI format 10 (FP16) / 87 (BGRA8) capture loop reported in #86. SDR capture now requests a stable BGRA8 format through `DuplicateOutput1`.

Clarification: fullscreen NeuralScreen intentionally uses `WDA_EXCLUDEFROMCAPTURE` to prevent recursive self-capture. Therefore Windows screenshots and ordinary OBS display capture do not include the NR layer. Use NeuralScreen's built-in screenshot, Spout2, or capture the one-window mode instead.

Still open / not implemented in v1.12:

- a persistent “always save screenshots” toggle and custom screenshot directory (the current folder only seeds the **Save As** dialog);
- dragging the fullscreen output betw...

v1.12.0 is now available: https://github.com/perseval-BLR/NeuralScreen/releases/tag/v1.12.0

This release does **not** claim to make the underlying RTX 4060 runtime `CreateFeature` access violation disappear. It does make this path safer and easier to diagnose: the requirements query and `CreateFeature` are now reported as separate stages, and a terminal `CreateFeature` fault goes directly to safe passthrough without unsafe retries or resource teardown.

Please retest with v1.12.0 and attach the full new `NeuralScreen.log`. If it still reports `CreateFeature raised`, we will keep this open as a runtime/driver compatibility issue rather than treating the earlier `0xBAD00012` requirements result as the cause.

Thanks for the log. v1.12.0 is now available: https://github.com/perseval-BLR/NeuralScreen/releases/tag/v1.12.0

It contains two fixes taken directly from this report:

- The flicker log showed SDR Desktop Duplication alternating between DXGI format 10 (FP16) and format 87 (BGRA8), rebuilding the capture bridge each time. v1.12 requests a stable BGRA8 format for SDR capture through `DuplicateOutput1`, so that format/rebuild loop should be gone on a 10-bit display.
- The 192 kHz WASAPI endpoint could not be opened by the AAC encoder, which aborted the MP4. Audio is now resampled to AAC at 48 kHz and probed before the real output file is opened. If AAC is unavailable, recording continues as video-only instead of leaving a broken or missing file.

Recording still saves automatically to the `recordings` folder; v1.12 does not add a video **Save As** dialog or custom recording destination yet.

Could you retest the 10-bit mouse-movement case and one short recording on v1.12.0? If either sti...