Search options (Page 1 of 9)

That's the answer then: on your card 4.6 is too heavy for a 4K vertical frame at x2, and 4.4 v2 fits.

I need to correct my Task Manager tip. With the same model and frame rates here (4.4 v2, 24 to 48 fps, no drops), nvidia-smi shows only 63-74% GPU utilization. So utilization doesn't reach 100% even at the limit, and skipped frames are the real test, which you already did.

One difference: my monitor is 1440p, not 4K. RIFE still processes the full 4K frame (SVP doesn't scale it to the window), only the final scaling in the renderer differs.

Your two screenshots come from different monitors. The horizontal video played on the XV272K (primary, 120 Hz) and the vertical one on the P272U PRO (the rotated one, running at 60 Hz). So the video and the monitor changed together, and the stats can't tell which one causes the drops. The vertical one also looks like it's falling behind before the renderer: 47.606 fps measured vs 47.952 target, while the horizontal one is at 47.984.

Could you swap them? Play the vertical video on the XV272K and the horizontal one on the P272U, with the same settings, and check Ctrl+J again. If the drops stay with the vertical video, it's the video/RIFE side. If they move to the P272U, it's the rotated monitor or its 60 Hz. Here the rotated monitor made no difference, but mine is 1440p and I tested with EVR-CP. You use MPC Video Renderer on a 4K screen, so your setup may behave differently.

Your screenshot shows the cause: your VapourSynth runs on Python 3.13, but the rife\pylib folder in my zip is built for Python 3.12, which is what SVP's own VapourSynth uses, so numpy can't load.

To fix it, either download the archive again and copy it over the same way:
https://drive.google.com/drive/folders/ … XX-S2JMOXD

or run this command, it installs the Python 3.13 versions of the packages TensorRT-RTX needs into your own Python (numpy, onnx and onnxconverter-common to convert the model to fp16, onnxruntime to prepare the v2 models):
d:\develop\python\python313\python.exe -m pip install numpy onnx onnxconverter-common onnxruntime
Your own packages are used before pylib, so nothing else needs changing. I reproduced both of your errors here with Python 3.13 + R79, and after this command RIFE worked with both the normal and v2 models.

About akarin.dll: the zip's rife folder is meant to be copied on top of SVP's rife folder (merge and overwrite), not to replace it. SVP's own rife folder already has akarin.dll, the models and the rest. If yours was replaced, restore it (SVP's maintenance tool or a reinstall), then copy the zip's rife folder over it again.

Tested it: no. RIFE (TRT and TRT-RTX, 1080p and 4K) runs at the same speed on R73 (what SVP ships), R79 and R80, within 0.6%. Plain svpflow (non-RIFE profile, 1080p and 4K) is the same on R73 and R79, within run to run noise. R80 dropped API 3 plugins, so SVP's svpflow, vstrt and akarin don't load on it at all; R79 is the newest SVP can use.

GPU load is 98%. I attached the benchmark I use (svp-rife-bench.zip), so you can
run the same test. It only needs an SVP 4 install: it uses SVP's own VSPipe and
a synthetic clip, so no video file, decoder or display is involved. The GPU is
the only thing working, and numbers are comparable between machines. Unpack, run
run_svp_rife_bench.cmd, details are in the README.

My numbers from that test (5090)

max output fps at x2, 2 GPU threads, best of 3 runs (a 24 fps video at x2 only
needs 48):

resolution  model     TensorRT   TensorRT-RTX   RTX vs TRT
1080p       4.6         312.8       312.5         same
1080p       4.6 v2      330.7       349.1         +5.6%
4K          4.6          71.9        74.8         +4.0%
4K          4.6 v2       77.6        83.6         +7.7%

v2 is also faster than v1: 6 to 8% on TensorRT, about 12% on TensorRT-RTX.

Archive updated: v2 models now work
https://drive.google.com/drive/folders/ … XX-S2JMOXD

The archive is updated to support the v2 models. That fixes the crash pituz hit.
Please re-download before testing. The v2 models are the "(v2)" entries in the
"AI model" list, for example "4.6 (v2)", and they show up there once their files
are in rife\models\rife_v2.

I also did what you suggested: generate.js is no longer patched. TensorRT-RTX is
now switched on with a "rife_rtx" user defined option. How to enable it:

1. Close SVP and any player, back up the SVP 4 folder.
2. Copy the "SVP 4" folder from the archive over "C:\Program Files (x86)\SVP 4"
(needs admin rights once).
3. Start SVP, then main menu -> Application settings -> "User defined options"
tab, fill in the fields below and press "Add option".

Title:          TensorRT-RTX   (any text, it is just the label)
Script name:    rife_rtx       (this exact name matters)
Option scope:   FRC profile
Allowed values: ON or OFF

4. Open your RIFE profile. At the bottom there is now a "User defined options"
section with a "TensorRT-RTX" row: set it to On. "Neural network engine" must
still be "NVIDIA TensorRT".
5. Play a video.

The second attachment is a screenshot of both places: the option being added on
the right, and the "TensorRT-RTX" row set to On in the profile on the left.

Off is the default and gives exactly stock SVP behaviour, so the option is safe
to leave in place.

What I bumped into: v2 models gave corrupted frames on TensorRT-RTX

Every interpolated frame had a checkerboard of blown pixels over the whole
picture. A speed test cannot show this, I only saw it when I compared the output
frames. I should have checked sooner.

What breaks:

* v2 model + TensorRT-RTX 1.6.1: corrupted
* v2 model + plain TensorRT: fine
* v1 model + either backend: fine

Why: the v2 models compute their warp grid from the tensor shape at runtime
(ops: Shape, Gather, Range, Expand, Where, Equal, ConstantOfShape). TensorRT-RTX
1.6.1 miscompiles that part of the graph. The v1 models do not have those ops.
It is not a precision problem: a pure fp32 engine shows the same corruption, and
the same fp16 graph is correct on plain TensorRT.

Fix (in the updated archive): constant-fold that part of the graph for the
padded resolution before building the engine. After that, TensorRT-RTX and
TensorRT output is bit identical. Folding takes 0.3 s, once per model and
resolution. The engine is already cached per resolution, so the folded model
sits next to it.

If you want to apply TensorRT-RTX, do this:

https://drive.google.com/drive/folders/ … XX-S2JMOXD

1. Close SVP and the player.
2. Make a backup of your "SVP 4" folder.
3. Copy the "SVP 4" folder from the archive on top of yours, overwrite all. Merge it, don't delete or replace your rife folder: SVP's own files in it are still needed.
4. Start SVP and add the switch once: main menu (the vertical dots) > Application settings > "User defined options" tab. Title: TensorRT-RTX (any text), Script name: rife_rtx (this exact name matters), Option scope: FRC profile, Allowed values: ON or OFF, then Add option.
5. Open the RIFE profile you use. At the bottom there is now a "TensorRT-RTX" row under "User defined options", set it to On. Engine stays "NVIDIA TensorRT".
6. Play a video.

Off gives stock SVP behaviour. The first time a model or a resolution is used
it prepares the engine by itself, after that it is cached.

If you use your own VapourSynth instead of SVP's, Python 3.12, 3.13 and 3.14 work as is. On any other Python version, also run this once in that Python:

<your python.exe> -m pip install numpy onnx onnxconverter-common onnxruntime

Numbers here, RTX 5090 Laptop, model 4.6, x2, 5 runs each:

                      TensorRT      TensorRT-RTX
1080p                261.8 fps       267.4 fps
4K                    60.0 fps        62.8 fps
engine build 1080p       103 s            12 s
engine build 4K          122 s            14 s
seek in mpv              0.9 s          0.45 s
runtime files          2565 MB          222 MB

The last row is what the two backends need on disk to build and run engines.
The archive does not remove anything from your install, it adds the RTX
runtime next to what you already have, but it shows what SVP could ship
instead of the old TensorRT 10 builder libraries.

Nevermind, asked Fable and it did some tests:

The setup: RTX 5090 Laptop GPU, SVP's bundled stack (TensorRT 10.8, vstrt.dll) vs TensorRT-RTX 1.6.1 (vs-mlrt v16.2.test1 vstrt_rtx.dll), RIFE 4.26 Heavy, fp16 + static shape + CUDA graph + 2 streams (SVP defaults), synthetic source through vspipe, so the numbers are max interpolation throughput and source fps does not change them. Numbers below are TensorRT vs TensorRT-RTX, median of 5 runs per config; the dashed marks in the chart show what a 23.976 movie needs at each multiplier:

1080p (1920x1088): x2 132.6 vs 135.4, x3 102.8 vs 101.1, x4 91.2 vs 90.3, x5 84.7 vs 84.1. Identical within run to run noise.
4K (3840x2176): x2 30.6 vs 33.0, x3 22.6 vs 24.4, x4 20.3 vs 21.4, x5 19.4 vs 20.4. TensorRT-RTX is 5-8% faster and won every single repetition.
For a 23.976 movie that means 4.26 Heavy is real-time at 1080p up to x3 (x4 barely misses, 91 vs 96 needed) and not real-time at 4K even at x2.

Engine build for 4.26 Heavy: 99 to 112 seconds on TensorRT 10.8 vs 0.5 seconds on TensorRT-RTX (about 7 seconds total setup including first engine load). Rebuilt engines from scratch twice per backend to confirm, it holds.

Also confirmed the weird-size crash is really fixed in 1.6.1: 1920x832, 1280x768 and 896x512 all build and render fine.

So: same speed at 1080p, faster at 4K, and the 100 second engine build basically disappears. Looks like a clear win. SVP's script generator already has the rife_rtx hook that loads vstrt_rtx.dll, it just needs the updated rife package (vstrt_rtx.dll + tensorrt_rtx DLLs + newer vsmlrt.py).

Install ChatGPT locally and tell it to play a video with SMplayer and fix any errors found.

It's nice to see AI models being able to improve existing apps with minimal knowledge. If you ever find some new features/fixes, please share them. Just add a clear explanation so it's description is easily searchable.
Good job!

I misread title.
Good answer btw!

If u add this in mpv.conf:
vf=d3d11vpp=scale=2.00:scaling-mode=nvidia

and you click the Nvidia Super Resolution checkbox for "Show status indicator on video" in NVIDIA App, it will show a small "RTX VSR" on top right side of mpv that Video Super Resolution is active:
https://gyazo.com/12f86d12f896441994a3421b2cc78af7

but changing DLSS models doesn't seem to do anything so no clue how we can use it at transcoding.