Vapoursynth Anime4KCPP
Anime4KCPP is a fast CNN upscaler for anime and cartoon content that enlarges by a free factor and can denoise lightly at the same time. It is one of the resizers of Vapoursynth Resizer and runs on the CPU, or on the GPU through OpenCL or CUDA.
Filtering → Vapoursynth → Frame → Resize

What it does
A small neural network enlarges the picture by Factor. The model decides the character of the result: neutral, mildly denoising, denoising and softening, or denoising and sharpening. When the upscaled size does not hit the output size exactly, a classic kernel adjusts it.
Requirements
- Windows: the plugin comes with Hybrid. Linux/macOS: offered only when the Anime4KCPP library is in the VapourSynth plugin folder.
- CPU processing works everywhere. OpenCL needs an OpenCL runtime; CUDA needs Windows and an NVIDIA GPU.
Settings
| Control | Parameter | Range / values | Hybrid default | Meaning |
|---|---|---|---|---|
| Factor | factor | 1.01 – 16.0 | 2.0 | Upscale factor above 1; need not be a whole number. |
| Processor | processor | cpu, opencl, cuda | cpu | Where the network runs; opencl and cuda are only listed when available. |
| Device | device | 0 – 6 | 0 | Device index for OpenCL/CUDA. |
| Model | model | 43 models: acnet-legacy-*, acnet-f8b4/b8/b18 (plain, -hdn, -box, -box-hdn), arnet-f8b8/b16/b32/b64 (plain, -hdn, -box, -box-hdn), artcnn-c4f16/c4f32 (plain, -dn, -ds), fsrcnnx-f8b4/f16b4 (plain, -distort-plus) | acnet-f8b8-hdn | Network used for the upscale, see Models. |
| Downsizer | – | point, rect, linear, cubic, lanczos, blackman, blackmanminlobe, spline36, spline64, spline144, spline, gauss, sinc | spline64 | Kernel that brings the result to the exact output size. |
Example
Default settings on a 10-bit YUV 4:2:0 source (640x356), resized to 1280x712:
clip = core.std.AddBorders(clip=clip, left=0, right=0, top=14, bottom=14) # add borders to achieve mod 32 (vsAnime4KCPP) - 640x384 # adjusting color space from YUV420P10 to YUV444PS for vsAnime4KCPP clip = core.resize.Bicubic(clip=clip, format=vs.YUV444PS) # resizing using Anime4KCPP # Zoomfactor: 2 clip = core.anime4kcpp.ACUpscale(clip, model="acnet-f8b8-hdn") # 1280x768 clip = core.std.Crop(clip=clip, left=0, right=0, top=28, bottom=28) # removing added borders from mod requirement (vsAnime4KCPP) - 1280x712
How Hybrid calls it
- Selected with Resize enabled and Resizer set to Anime4KCPP (see Vapoursynth Resizer).
- Accepted input: YUV 4:4:4, 4:2:2 and 4:2:0 in 8 bit, 16 bit and 32-bit float, and gray in 8 bit, 16 bit and float; other formats are converted first. Width and height are brought to mod 32 before the call.
- Hybrid expects the output to be input × factor (whole pixels). If that differs from the target size, fmtconv resizes the result with the Downsizer kernel to the exact size, also when that means shrinking. With separately filtered fields the target height is halved.
- The model is always written; factor, processor and device only when they differ from the defaults (which are also the plugin defaults).
Background
Models
The bundled Anime4KCPP 3.2 describes its models like this:
- acnet-legacy-gan enhances detail; acnet-legacy-hdn0 to hdn3 are the older lightweight CNNs with moderate to extreme denoising.
- acnet-f8b… are lightweight VGG-style networks, arnet-f8b… lightweight ResNet-style networks; the larger the number after "b", the bigger and slower the network. Plain models are trained to be neutral, -hdn to denoise mildly. -box variants were trained with box degradation, which restores lines better but can look slightly blurry.
- artcnn-… are Artoriuz's ArtCNN models: neutral, -dn denoises and softens, -ds denoises and sharpens.
- fsrcnnx-… are FSRCNNX models (igv) that denoise slightly; the -distort-plus versions (nessotrin) denoise strongly.
Anime4KCPP
Anime4KCPP (TianZerL) is a C++ implementation of anime upscaling with its own ACNet networks. The same library offers filters for VapourSynth and AviSynth+; the VapourSynth function also accepts factors just above 1.
Sources
- Anime4KCPP
- ArtCNN
- FSRCNNX
- Hybrid's tooltips