Motion Estimation wobbly

[Petter005]

I have a RTX 3070 and if i select NVOFA everything starts to look like gelatin and some pixels move around, if i use CPU DIS everything looks normal, the NR and the FG works (with DLSSG for SM86 mod) but i noticed that difference in the 2 options we have for motion estimation

[perseval-BLR]

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, goes to the network unfiltered. "Gelatin and pixels moving around" is what that looks like: vectors that explain nothing still move the picture.

There is a length cut in the expand shader (vectors below 0.5 px are zeroed), but that is a length test only - it cannot tell a real vector from a long wrong one, which is the case the static hypothesis catches.

What I need to make this actionable

Your NeuralScreen.log. The worker logs one of:

[nvofa] active: 320x180, grid=4, cost=8-bit; driver D3D12 API
[nvofa] unavailable: <operation> (<status>) <detail>

That tells me whether the driver took the path we expect on your 3070 (grid size, and whether the caps query agreed), or whether it fell back and you were looking at something else. If it is easy, the before/after wipe with NVOFA on, over moving content, would help too.

Meanwhile

Settings -> Capture -> Motion estimation -> CPU DIS is the way back to what you had; it needs a worker restart, which the program does by itself. Nothing else has to change.

I am not going to promise a fix or a date. The honest statement of the problem: the driver's motion field needs the same two tests DIS gets, or at least the static hypothesis, and that work is on the worker side. Whether it lands as validation or as a different default is a decision I have not made.

Also worth saying plainly: NVOFA became the default in v1.13.1. If your log shows it active and healthy on a 3070 while the picture looks like this, that is an argument against the default for that generation, and I would rather know it now.

[Petter005]

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

[perseval-BLR]

@Petter005 - yes, a log with the different settings is exactly what would settle it, and thank you for offering.

What makes the wobbly motion readable is the backend line and the numbers next to it, so if you can, run it twice on the same scene - once with motion estimation on the GPU backend (NVOFA) and once on the CPU one - and send both logs. The pair I look at is the [nvofa] lines against the [perf] guides timing: if the GPU path is refused it falls back to the CPU silently and the guide timing jumps, which reads as wobble on screen rather than as an error. Same scene and the same camera movement matter more than a long session.

GitHub is fine, and better than Discord for this: the next person with the symptom searches the issues, not a chat server. If it is the fallback there is a line that says so and I will quote it back; if it is something else, the two logs still tell me which half of the pipeline is doing it.

Current release: https://github.com/perseval-BLR/NeuralScreen/releases/tag/v2.0.1

[Petter005]

NeuralScreen cpu DIS.log
NeuralScreen NVOFA.log

2026-09-23.19-34-40.mp4