GLSL (v4.0.1 Stable)

Installation Instructions for GLSL/MPV (v4.0.1)

What's new:

• High quality Restore_Soft shaders (5 variants) to address different types of artifacts. Soft variants significantly reduces ringing, aliasing and other artifacts at the expense of less sharpness.
• Simplified default shaders usage without compromising on customizability.
• Complicated use cases (>12 different shader combinations) were simplified and grouped into 3 modes (A, B and C). (See installation instructions)

Major Changes:

• Better installation/customization guide.
• Light variants of Restore shaders are replaced by Soft variants.
• Improved line restoration shaders.

Minor Changes:

• Restore_Moderate shaders are renamed to simply Restore.
• Added AutoDownscalePre_x2 shader for better compatibility with 1440p/1080p monitors.

GLSL (v4.0.0 Release Candidate)

Installation Instructions for GLSL/MPV (v4.0)

What's new:

• Every change from v3.2 is included.
• Ten variants for line reconstruction shaders. 2 Strengths with 5 network sizes (S/M/L/VL/UL).
• Improved line reconstruction shaders.
• Reduced unwanted ringing, aliasing and banding artifacts for line reconstruction shaders.
• Improved text legibility for line reconstruction shaders.

Major Changes:

• All shaders now work on RGB space instead of Luma.
• This version breaks compatibility with all versions before v3.2! Mixing with older shaders might cause issues.
• Removed RA, Upscale+Deblur CNN and Deblur CNN shaders, as they are superseded by v4.x Restore shaders.
• Removed Heavy variant for line reconstruction. The Moderate variant has been tweaked for better reconstruction of low quality images.

Minor Changes:

• Standardized network/model size convention. Each increase in size doubles the amount of computation required.
• Standardized shader naming convention. All shader names now take the form Anime4K-[Type-Subtype]-[Strength]-[Quality/Speed].
• Line reconstruction shaders renamed to a shorter name: Restore.

GLSL (v3.2 Stable)

What's new:

• Improved CNN upscale and denoise shaders.
• Ten variants for upscale shaders. 2 types (None/Denoise) with 5 network sizes (S/M/L/VL/UL).

Major Changes:

• Updated shaders for compatibility with future v4.x shaders.
• All shaders now work on RGB space instead of Luma.
• This version breaks compatibility with all previous versions! Mixing with older shaders might cause issues.
• Removed RA, Upscale+Deblur CNN and Deblur CNN shaders, as they are superseded by v4.x Restore shaders.

Minor Changes:

• Standardized network/model size convention. Each increase in size doubles the amount of computation required.
• Standardized shader naming convention. All shader names now take the form Anime4K-[Type-Subtype]-[Strength]-[Quality/Speed].
• Non-CNN Upscale and Deblur shaders are now deprecated. They might be removed in future versions.
• Small tweaks for non-CNN shaders improving quality for low resolution inputs.
• Renamed DarkLines and ThinLines shaders to a shorter name: Darken and Thin.

GLSL (v4.0.0 Alpha) Experimental

Huge quality improvement on ITW (In the wild) anime. Previous super resolution methods were suited for artificially downscaled anime, while anime found in the wild exhibit a large amounts of variance. We propose and develop a method to correct this distribution shift for already existing CNN upscalers and that can be applied to non-Anime4K upscalers (The upscaling shaders bundled in this version are the exact same as v3.1). The line reconstruction algorithm will be explained in a future informal paper.

Some small issues such as occasional ghosting or ringing will need to be fixed before official release.

The three versions of line reconstruction shaders can be used as follows in mpv's input.conf:
CTRL+2 works on almost all 1080p anime, if blurriness and aliasing is noticeable, use the heavier version (CTRL+3), if ghosting and ringing is visible, use the lighter version (CTRL+1).
For downscaled 720/480p anime, start with CTRL+1, for native standard definition anime, start with CTRL+2.

input.conf for 4K screens (screen size, not anime size)

CTRL+1 change-list glsl-shaders set "~~/shaders/Anime4K_Line_Reconstruction_Light_L.glsl;~~/shaders/Anime4K_Upscale_CNN_L_x2_Denoise.glsl;~~/shaders/Anime4K_Auto_Downscale_Pre_x4.glsl;~~/shaders/Anime4K_Line_Reconstruction_Heavy_L.glsl;~~/shaders/Anime4K_Upscale_CNN_M_x2_Denoise.glsl;~~/shaders/Anime4K_DeRing.glsl;"
CTRL+2 change-list glsl-shaders set "~~/shaders/Anime4K_Line_Reconstruction_Medium_L.glsl;~~/shaders/Anime4K_Upscale_CNN_L_x2_Denoise.glsl;~~/shaders/Anime4K_Auto_Downscale_Pre_x4.glsl;~~/shaders/Anime4K_Line_Reconstruction_Heavy_L.glsl;~~/shaders/Anime4K_Upscale_CNN_M_x2_Denoise.glsl;~~/shaders/Anime4K_DeRing.glsl;"
CTRL+3 change-list glsl-shaders set "~~/shaders/Anime4K_Line_Reconstruction_Heavy_L.glsl;~~/shaders/Anime4K_Upscale_CNN_L_x2_Denoise.glsl;~~/shaders/Anime4K_Auto_Downscale_Pre_x4.glsl;~~/shaders/Anime4K_Line_Reconstruction_Medium_L.glsl;~~/shaders/Anime4K_Upscale_CNN_M_x2_Denoise.glsl;~~/shaders/Anime4K_DeRing.glsl;"
CTRL+0 change-list glsl-shaders clr ""

input.conf for 1080p screens (screen size, not anime size)

CTRL+1 change-list glsl-shaders set "~~/shaders/Anime4K_Line_Reconstruction_Light_L.glsl;~~/shaders/Anime4K_Upscale_CNN_L_x2_Denoise.glsl;~~/shaders/Anime4K_Auto_Downscale_Pre_x4.glsl;~~/shaders/Anime4K_Line_Reconstruction_Medium_L.glsl;~~/shaders/Anime4K_Upscale_CNN_M_x2_Denoise.glsl;~~/shaders/Anime4K_DeRing.glsl;"
CTRL+2 change-list glsl-shaders set "~~/shaders/Anime4K_Line_Reconstruction_Medium_L.glsl;~~/shaders/Anime4K_Upscale_CNN_L_x2_Denoise.glsl;~~/shaders/Anime4K_Auto_Downscale_Pre_x4.glsl;~~/shaders/Anime4K_Line_Reconstruction_Light_L.glsl;~~/shaders/Anime4K_Upscale_CNN_M_x2_Denoise.glsl;~~/shaders/Anime4K_DeRing.glsl;"
CTRL+3 change-list glsl-shaders set "~~/shaders/Anime4K_Line_Reconstruction_Heavy_L.glsl;~~/shaders/Anime4K_Upscale_CNN_L_x2_Denoise.glsl;~~/shaders/Anime4K_Auto_Downscale_Pre_x4.glsl;~~/shaders/Anime4K_Line_Reconstruction_Light_L.glsl;~~/shaders/Anime4K_Upscale_CNN_M_x2_Denoise.glsl;~~/shaders/Anime4K_DeRing.glsl;"
CTRL+0 change-list glsl-shaders clr ""

Note: This version is only compatible with v3.1 and older shaders. Mixing with newer shaders might cause issues.

GLSL (v3.1 Stable)

Installation Instructions for GLSL/MPV (v3.1)

What's new:

• A new wiki with information about every shader!
• A standalone DTD shader tweaked for fast perceptual upscaling.
• Two more variants of bilateral filters.

Minor changes:

• Removed the version number from the file name. Users will not have to update the config file anymore for future updates.

Bug fixes:

• Fixed texture size issues in RA, Upscale_CNN, and Upscale_CNN_Deblur shaders that caused faulty processing when stacked one after the other.
• RA shaders now correctly ignore chroma channels. They used to wrongly process chroma channels.

GLSL (v3.0 Stable)

Installation Instructions for GLSL/MPV (v3.0)

What's new:

• The shader is now broken into modular parts, allowing extensive customization.
• Better quality CNN line art super-resolution shaders. (Faster and higher PSNR compared to all NGU and FSRCNNX variants.)
• A large amount of shaders designed to improve visual/perceptual quality.
• Shaders designed to reduce noise and artifacts.

Minor changes:

• The SR CNNs were trained on MAE loss with the SYNLA dataset, instead of perceptual loss and images randomly collected in the wild in v2.1. This significantly improves accuracy for recovering detail in low resolution videos at the cost of some perceptual quality. However this can be mitigated by using the aforementioned shaders.

GLSL (v2.1a Hotfix)

Installation Instructions for GLSL/MPV Version

Hotfix: Algorithm now performs downscaling before applying the x4 prescaler (rather than after). 720p scaled to x4 is bigger than 4K, which is unnecessary and slows the algorithm down.

Added a L version that contains a larger prescaler. It offers a significant improvement to image quality, especially to low resolution anime. (eg. 480p) It has 2741 parameters, which is still smaller than FSRCNNX 8. If your GPU can run FSRCNNX, it should be fine.
However, since the prescaler was trained using perceptual loss, some banding might occur on very old/blurry anime.
It is also slower. (+1ms compared to non-L version for 1080p -> 4K upscaling)
Use at your own discretion.

GLSL (v2.1 Stable)

Installation Instructions for GLSL/MPV Version

Integrated a pre-scaler in Anime4K. It is a tiny ConvNet (916 parameters, 3 hidden layers of 4 feature maps) trained with the differentiable Anime4K algorithm and other techniques that will be published soon. For comparison, FSRCNNX has more than 3000 parameters (depending on the version) and waifu2x has more than 1 million.

This pre-scaler gives a major quality improvement to lower resolution anime, especially for 720p and 480p.
480p -> 1080p, 480p -> 4K and 720p -> 4K are now officially supported.
Texture loss and aliasing issues are again reduced.

The pre-scaler uses 1.5ms for 1080p -> 4K upscaling, bringing the total to 4ms on a Vega 64 GPU.

GLSL (v2.0 Release Candidate 5) Experimental

Installation Instructions for GLSL/MPV Version

Replaced the Bilateral Mean Filter with a "Regularized" Bilateral Mode Filter. The mode is more robust than the mean or median, and introduces the least amount of blur.

The mode filter needs to be "regularized" since the mode of a set is not defined when the number of data points is finite and contains unique data (all data points are different). This undefined case happens extremely often in images when using a small kernel (eg. 3x3), since there are 256 possible luma values, but only 9 samples. Often all 9 pixels will have different values, and rigging up a continuous fallback to median/mean that does not introduce discontinuities is not trivial.

However, we can use Kernel Density Estimation (KDE) as a continuous approximation of the discrete probability color histogram.

GLSL (v2.0 Release Candidate 4) Experimental

Installation Instructions for GLSL/MPV Version

Added a bilateral filter to Anime4K. This reduces noise and blocking on highly compressed video (mostly found on the web). Speed is not affected.

Anime4K now enhances edges even in cases where upscaling is not necessary, as seen in this comparison.

On a more personal note, I now feel as if ML based techniques for denoising line art are just mostly re-learning traditional computer graphics/image processing techniques, + ε. The small epsilon leads to a slightly better image quality, but at what cost to speed? (Of course this does not apply for extreme cases such as low resolution upscaling, denosing with a negative signal to noise ratio or image style transfer, since ConvNets can "generate" missing information).