diff --git a/src/github.com/matrix-org/dendrite/mediaapi/README.md b/src/github.com/matrix-org/dendrite/mediaapi/README.md
index 8a32263bb..8d6cc6270 100644
--- a/src/github.com/matrix-org/dendrite/mediaapi/README.md
+++ b/src/github.com/matrix-org/dendrite/mediaapi/README.md
@@ -4,6 +4,24 @@ 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.
 
-Alternatively one can use `gb 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.
+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:
+
+![](nfnt-96x96-crop.jpg)
+
+### bimg (uses libvips C library)
+
+Alternatively one can use `gb 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:
+
+![](bimg-96x96-crop.jpg)
diff --git a/src/github.com/matrix-org/dendrite/mediaapi/bimg-96x96-crop.jpg b/src/github.com/matrix-org/dendrite/mediaapi/bimg-96x96-crop.jpg
new file mode 100644
index 000000000..f6521893a
Binary files /dev/null and b/src/github.com/matrix-org/dendrite/mediaapi/bimg-96x96-crop.jpg differ
diff --git a/src/github.com/matrix-org/dendrite/mediaapi/nfnt-96x96-crop.jpg b/src/github.com/matrix-org/dendrite/mediaapi/nfnt-96x96-crop.jpg
new file mode 100644
index 000000000..1e424cd8b
Binary files /dev/null and b/src/github.com/matrix-org/dendrite/mediaapi/nfnt-96x96-crop.jpg differ