Upscaling

There are many available upscaling (interpolation) techniques in mpv, of which this article aims to provide some comparison screenshots.

Note: This article has mostly been superseded by this repository, which catalogs and compares more upscalers and test files than this page, and will also be kept more up-to-date.

Methodology

The source image for all of these is imagemagic's built-in “rose” test pattern, which can be obtained like this:

convert rose: rose.png

Each of the screenshots was obtained like this:

mpv --no-config --pause rose.png --geometry=560x368 -vo opengl-hq:dither-depth=8:scale=SCALER

The resulting output was captured using imagemagick:

import -depth 8 SCALER.png

Results

scale=bilinear

scale=lanczos

scale=mitchell

scale=spline16

scale=spline36

scale=spline64

scale=ewa_lanczos

scale=ewa_lanczos:scale-radius=3.2383154841662362

scale=ewa_lanczossharp

scale=ewa_lanczossoft

scale=robidoux

scale=robidouxsharp

scale=ewa_ginseng

scale=ewa_hanning

scale=bicubic

scale=gaussian

scale=catmull_rom

scale=nearest

scale=oversample

scale=sharpen3

scale=sharpen5

scale=sinc:scale-radius=16

Antiringing

Some of the upscalers support antiringing, which is a naive technique that attempts to reduce ringing in some cases, but can also introduce artifacts. The antiringing algorithm in mpv is not very sophisticated.

scale=spline36:scale-antiring=1

scale=lanczos:scale-antiring=1

scale=ewa_lanczossharp:scale-antiring=1

Other upscaling methods

There are also ways to upscale images other than by using the scale option, such as using a vapoursynth script. (Note that using vapoursynth on still images in mpv is somewhat tricky, and working with RGB colors is also difficult. To produce these images, the source PNG was split into three separate grayscale images (for R, G and B) and merged into a movie clip of length 3. Each frame of the output was captured and converted back to an RGB image)

core.nnedi3.nnedi3_rpow2(video_in, 8).set_output()