Skip to content

Super-Resolution Convolutional Neural Networks as GLSL shaders for mpv

License

Notifications You must be signed in to change notification settings

Artoriuz/ArtCNN

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ArtCNN

Overview

ArtCNN is a collection of SISR CNNs optimised for anime content.

Two distinct architectures are currently offered:

  • R: Bigger model aimed mostly at encoding tasks. On top of having more filters per convolution layer, the model was also made much deeper with the help of residual blocks and short-skip connections. Offered in the ONNX format.
  • C: Original ArtCNN models optimised mostly for speed. The architecture consists of a series of convolution layers aided by a single long-skip connection. Offered as GLSL shaders to be used in real-time on mpv.

The R architecture is offered in 2 sizes:

  • R16F96: This has 16 residual blocks and 96 filters per convolution layer. Should generally give you the best reconstruction quality.
  • R8F64: This has 8 residual blocks and 64 filters per convolution layer. An attempt at balancing quality and performance.

The C architecture is offered in 3 sizes:

  • C4F32: This has 4 internal convolution layers with 32 filters each. You need a relatively decent GPU to run this well.
  • C4F16: This has 4 internal convolution layers with 16 filters each. You should be able to run this on most modern GPUs.
  • C4F8: This has 4 internal convolution layers with 8 filters each. You should probably only use this on very slow systems.

Regarding the suffixes:

  • Models without any suffixes are the baseline. These are neutral luma doublers.
  • CMP variants are compute shaders. These are generally much faster on modern GPUs, specially on Vulkan. You should always try to run these variants first, and only fallback to the fragment shaders if you encounter problems and/or performance issues.
  • DS variants are trained to denoise and sharpen, which is usually useful for most web sources.
  • Chroma variants are trained to reconstruct chroma. These are intended to be used on 4:2:0 content and will not work as intended in any other scenario (which means you can't use them after luma doublers).

You may occasionaly find some models under the "Experiments" directory. This is meant to serve as a testing grounds.

mpv Instructions

Add something like this to your mpv config:

glsl-shader="path/to/shader/ArtCNN_C4F16_DS_CMP.glsl"

VapourSynth Instructions

ArtCNN is natively supported by vs-mlrt. Please follow the instructions found there.

Alternatively, can also run the GLSL shaders with vs-placebo.

About

Super-Resolution Convolutional Neural Networks as GLSL shaders for mpv

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages