v3.2.0

The ACNet and ARNet models have been redesigned and retrained. The previous ACNet model is now referred to as ACNet-Legacy.

The ACNet-Legacy model includes 5 variants, from hdn0 to hdn3, with progressively stronger denoising capabilities. Additionally, a gan variant has been introduced to enhance details.

ACNet and ARNet offer models of various sizes, which can be selected based on your hardware configuration. Each size provides 4 variants:

• no suffix: trained to be neutral.
• hdn: trained to mildly denoise.
• box: trained to be neutral through box degradation, which is better for line restoration, but may appear slightly blurry.
• box-hdn: trained to mildly denoise based on the box variant.

All shaders have been reimplemented with a new approach, adding support for mpv-android (fix #7).

If you are interested in the training of ACNet and ARNet, or the scripts for exporting them to GLSL, please check out the training code.

Models of Anime4KCPP v3.2.0:

models	parameters	description
acnet-legacy-gan	4784	Lightweight CNN, detail enhancement.
acnet-legacy-hdn0	4784	Lightweight CNN, moderate denoising.
acnet-legacy-hdn1	4784	Lightweight CNN, strong denoising.
acnet-legacy-hdn2	4784	Lightweight CNN, aggressive denoising.
acnet-legacy-hdn3	4784	Lightweight CNN, extreme denoising.
acnet-f8b4	2748	Lightweight VGG-style network, trained to be neutral.
acnet-f8b4-hdn	2748	Lightweight VGG-style network, trained to mildly denoise.
acnet-f8b4-box	2748	Lightweight VGG-style network, trained to be neutral through box degradation which is better for line restoration, but it may appear slightly blurry.
acnet-f8b4-box-hdn	2748	Lightweight VGG-style network, trained to mildly denoise based on the box variant.
acnet-f8b8	5116	Lightweight VGG-style network, trained to be neutral.
acnet-f8b8-hdn	5116	Lightweight VGG-style network, trained to mildly denoise.
acnet-f8b8-box	5116	Lightweight VGG-style network, trained to be neutral through box degradation which is better for line restoration, but it may appear slightly blurry.
acnet-f8b8-box-hdn	5116	Lightweight VGG-style network, trained to mildly denoise based on the box variant.
acnet-f8b18	11036	Lightweight VGG-style network, trained to be neutral.
acnet-f8b18-hdn	11036	Lightweight VGG-style network, trained to mildly denoise.
acnet-f8b18-box	11036	Lightweight VGG-style network, trained to be neutral through box degradation which is better for line restoration, but it may appear slightly blurry.
acnet-f8b18-box-hdn	11036	Lightweight VGG-style network, trained to mildly denoise based on the box variant.
arnet-f8b8	9860	Lightweight ResNet-style network, trained to be neutral.
arnet-f8b8-hdn	9860	Lightweight ResNet-style network, trained to mildly denoise.
arnet-f8b8-box	9860	Lightweight ResNet-style network, trained to be neutral through box degradation which is better for line restoration, but it may appear slightly blurry.
arnet-f8b8-box-hdn	9860	Lightweight ResNet-style network, trained to mildly denoise based on the box variant.
arnet-f8b16	19268	Lightweight ResNet-style network, trained to be neutral.
arnet-f8b16-hdn	19268	Lightweight ResNet-style network, trained to mildly denoise.
arnet-f8b16-box	19268	Lightweight ResNet-style network, trained to be neutral through box degradation which is better for line restoration, but it may appear slightly blurry.
arnet-f8b16-box-hdn	19268	Lightweight ResNet-style network, trained to mildly denoise based on the box variant.
arnet-f8b32	38084	Lightweight ResNet-style network, trained to be neutral.
arnet-f8b32-hdn	38084	Lightweight ResNet-style network, trained to mildly denoise.
arnet-f8b32-box	38084	Lightweight ResNet-style network, trained to be neutral through box degradation which is better for line restoration, but it may appear slightly blurry.
arnet-f8b32-box-hdn	38084	Lightweight ResNet-style network, trained to mildly denoise based on the box variant.
arnet-f8b64	75716	Lightweight ResNet-style network, trained to be neutral.
arnet-f8b64-hdn	75716	Lightweight ResNet-style network, trained to mildly denoise.
arnet-f8b64-box	75716	Lightweight ResNet-style network, trained to be neutral through box degradation which is better for line restoration, but it may appear slightly blurry.
arnet-f8b64-box-hdn	75716	Lightweight ResNet-style network, trained to mildly denoise based on the box variant.

v3.1.0

Models of Anime4KCPP v3.1.0:

models	description	parameters
acnet-gan	Lightweight CNN, detail enhancement	5k
acnet-hdn0	Lightweight CNN, mild denoising	5k
acnet-hdn1	Lightweight CNN, moderate denoising	5k
acnet-hdn2	Lightweight CNN, heavy denoising	5k
acnet-hdn3	Lightweight CNN, extreme denoising	5k
arnet-4-hdn	Lightweight ResNet, mild denoising	5k
arnet-4-le	Lightweight ResNet, line enhancing	5k
arnet-b8-hdn	Lightweight ResNet, mild denoising	10k
arnet-b8-le	Lightweight ResNet, line enhancing	10k
arnet-b16-hdn	Lightweight ResNet, mild denoising	19k
arnet-b16-le	Lightweight ResNet, line enhancing	19k
arnet-b24-hdn	Lightweight ResNet, mild denoising	28k
arnet-b24-le	Lightweight ResNet, line enhancing	28k
arnet-b32-hdn	Lightweight ResNet, mild denoising	37k
arnet-b32-le	Lightweight ResNet, line enhancing	37k
arnet-b48-hdn	Lightweight ResNet, mild denoising	55k
arnet-b48-le	Lightweight ResNet, line enhancing	55k
arnet-b64-hdn	Lightweight ResNet, mild denoising	74k
arnet-b64-le	Lightweight ResNet, line enhancing	74k

The ACNet model remains the same as before, but it has been uniformly updated with a new implementation, which is theoretically faster. The ACNet model includes 5 variants, from HDN0 to HDN3, with progressively stronger denoising capabilities. Additionally, a GAN variant has been introduced to enhance details.

The ARNet series includes 7 models of different sizes, ranging from B4 to B64, with increasing depth and parameter counts, resulting in improved quality. Each size offers 2 variants: the LE variant enhances lines but may alter the visual style. If you prefer not to have this effect, you can use the HDN variant. The HDN variant of ARNet is designed to preserve the original image’s look and feel, with only mild denoising (even lighter than ACNet HDN0).

ACNet GLSL

ACNet glsl shaders for MPV based player.

HDN shaders can better denoise, HDN level is from 1 to 3, higher for better denoising but may cause blur and lack of detail.