ArtCNN is a collection of simple SISR CNNs aimed at anime content.
Two distinct architectures are currently offered:
R: Bigger model aimed mostly at non real-time tasks like rescaling. 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 only in the ONNX format.C: Original ArtCNN models optimised mostly for speed. These should only be used for real-time tasks like video playback. The architecture consists of a series of convolution layers aided by a single long-skip connection. Offered in the ONNX format and as GLSL shaders.
4 sizes are currently offered:
| Model | Architecture | Residual Blocks/Layers | Filters | Parameter Count | Recommended Usage |
|---|---|---|---|---|---|
R16F96 |
R | 16 | 96 | ~4m | Highest-quality reconstruction for non real-time tasks |
R8F64 |
R | 8 | 64 | ~926k | Balanced option for non real-time tasks |
C4F32 |
C | 4 | 32 | ~48k | Real-time tasks if hardware allows |
C4F16 |
C | 4 | 16 | ~12k | Lightweight option for real-time tasks |
Regarding the suffixes:
- Models without any suffixes are the baselines. These are neutral luma doublers.
DSvariants are trained to denoise and sharpen, which is usually useful for most web sources.Chromavariants are trained to reconstruct chroma. These are intended to be used on 4:2:0 content and chroma must be upscaled with bilinear.
You may occasionaly find some experimental models under the Experiments directory.
Add something like this to your mpv config:
vo=gpu-next
glsl-shader="path/to/shader/ArtCNN_C4F32_DS.glsl"
ArtCNN is supported by vs-mlrt, with an easy-to-use wrapper in vs-jetpack. Please follow the instructions found there.
Alternatively, can also run the GLSL shaders with vs-placebo.
