dendrite/mediaapi
hacktivista 1ec5b0709b
Support WebP based thumbnails (#2033)
Co-authored-by: Felix Freeman <libsys@hacktivista.org>
2021-11-02 10:15:13 +00:00
..
fileutils - Removed double imports (#1989) 2021-09-08 17:31:03 +01:00
routing Set MaxFileSizeBytes <= 0 to unlimited (#1876) 2021-07-19 17:58:51 +01:00
storage Only include go-sqlite3 on the relevant binaries (#1900) 2021-07-20 11:18:14 +01:00
thumbnailer Support WebP based thumbnails (#2033) 2021-11-02 10:15:13 +00:00
types Top-level setup package (#1605) 2020-12-02 17:41:00 +00:00
bimg-96x96-crop.jpg use go module for dependencies (#594) 2019-05-21 21:56:55 +01:00
mediaapi.go Top-level setup package (#1605) 2020-12-02 17:41:00 +00:00
nfnt-96x96-crop.jpg use go module for dependencies (#594) 2019-05-21 21:56:55 +01:00
README.md Update Docker files and remove references to gb (#703) 2019-05-22 10:44:25 +01:00

Media API

This server is responsible for serving /media requests as per:

http://matrix.org/docs/spec/client_server/r0.2.0.html#id43

Scaling libraries

nfnt/resize (default)

Thumbnailing uses https://github.com/nfnt/resize by default which is a pure golang image scaling library relying on image codecs from the standard library. It is ISC-licensed.

It is multi-threaded and uses Lanczos3 so produces sharp images. Using Lanczos3 all the way makes it slower than some other approaches like bimg. (~845ms in total for pre-generating 32x32-crop, 96x96-crop, 320x240-scale, 640x480-scale and 800x600-scale from a given JPEG image on a given machine.)

See the sample below for image quality with nfnt/resize:

bimg (uses libvips C library)

Alternatively one can use go build -tags bimg to use bimg from https://github.com/h2non/bimg (MIT-licensed) which uses libvips from https://github.com/jcupitt/libvips (LGPL v2.1+ -licensed). libvips is a C library and must be installed/built separately. See the github page for details. Also note that libvips in turn has dependencies with a selection of FOSS licenses.

bimg and libvips have significantly better performance than nfnt/resize but produce slightly less-sharp images. bimg uses a box filter for downscaling to within about 200% of the target scale and then uses Lanczos3 for the last bit. This is a much faster approach but comes at the expense of sharpness. (~295ms in total for pre-generating 32x32-crop, 96x96-crop, 320x240-scale, 640x480-scale and 800x600-scale from a given JPEG image on a given machine.)

See the sample below for image quality with bimg: