Ristellise / VapourSynth-DPIR

VapourSynth DPIR Implementation

VapourSynth-DPIR

Ported from https://github.com/cszn/DPIR, https://github.com/HolyWu/vs-dpir

Still in beta.

If you only work on CPU, you can set parallel and have a speed improvement at the cost of maxed RAM (probably) If you're on GPU you will see the same speed.

Dependencies

• Torch (Tested with 1.9)
  • Cuda 10 (Tested on Cuda 11.1)
• VapourSynth (duh)

VapourSynth DPIR Implementation

Usage

  core.DPIR.Deblock(clip clip, [float strength, int gpu, int parallel, int no_cache])

  core.DPIR.Denoise(clip clip, [float strength, int gpu, int parallel, int no_cache])

• clip: Clip to process. Only RGBS format is supported.
• strength: Strenght of the function.
• gpu:
  • -1: Auto
  • 0: Cpu
  • 1: Cuda device
  • >1: Intended for multi GPU, will select device with index gpu - 1
• parallel: if to do parallel processing or not. 0 is recommended if you have a small amount of VRAM.
• no_cache: Disable cuda cache.

Compilation

I prefer doing it in VSCode, but

cmake --build ./build --config Release --target vsdpir -j 26 --

Keep in mind you'll have to set VS_LIB_PATH, TORCH_LIB_PATH, BOOST_LIB_PATH manually.

"Todo List"

Implement refield tensor in c++ (pain), so it can work on non mod-8 clips.

Implement a semaphore with cuda api to limit concurrent threads so it won't try to allocate absurd amounts of VRAM.

Add Deblur/Demosaick/SuperResolution.

If possible, OpenCL or ROCm optimization.