TensorRT Guide

April 8, 2025 · View on GitHub

This guide helps you use TSCUNet with TensorRT for accelerated video upscaling.

Current Limitations:

  • Static shapes required, must be multiples of 64 in each dimension
  • Input videos must either:
    • Be padded to match engine dimensions (e.g., 720x480 → 768x512)
    • Fit exactly to required dimensions

Setup Process

  1. Set up VapourSynth following pifroggi's guide
  2. Download and extract vsmlrt-windows-x64-tensorrt.[version].7z from vs-mlrt releases to your vs-plugins directory
  3. Get the model:
    • Download pre-converted ONNX from releases, or
    • Convert your own using convert_to_onnx.py (see script for detailed options)

Usage

  1. Build TensorRT engine using trtexec:

    FP32:

    trtexec --onnx="tscunet_fp32.onnx" --optShapes=input:1x15x512x768 --saveEngine=tscunet_fp32.engine --builderOptimizationLevel=5 --useCudaGraph --tacticSources=+CUDNN,-CUBLAS,-CUBLAS_LT

    FP16:

    trtexec --onnx="tscunet_fp16.onnx" --fp16 --optShapes=input:1x15x512x768 --inputIOFormats=fp16:chw --outputIOFormats=fp16:chw --saveEngine=tscunet_fp16.engine --builderOptimizationLevel=5 --useCudaGraph --tacticSources=+CUDNN,-CUBLAS,-CUBLAS_LT
    • You'll want to change the shape in the --optShapes=input:1x15x512x768 section depending on the resolution of your input video
    • Please note that the shape has to be a multiple of 64. So if your input video is 720x540, you'd want to use --optShapes=input:1x15x512x768

  2. Copy vapoursynth_script.py to your VapourSynth directory, then configure it with your video path and engine path

  3. Open a Command Prompt window (NOT POWERSHELL) in your VapourSynth directory, then run a command like this. Customize the encoder settings as you wish:

vspipe -c y4m ".\vapoursynth_script.vpy" - | ffmpeg -i - -c:v hevc_nvenc -qp 0 -preset p5 -tune lossless "output.mkv"