Expanded my emoting wife collection.
My review of dithering algorithms available in ffmpeg
>inb4 who asked
Dither algorithm summary (Starting with a command that would produce a bilinearly-downscaled 3.58 MB GIF with dither=none option)
- Bayer 3.91 MB Arbitrary Rating: shit/10
looks like shit. Seems to be dithering using the "microwave mesh grating" method - i.e. it overlays a very obvious and stupid-looking static pattern over the entirety of the animation.
- Heckbert 6.15 MB Arbitrary Rating: 2/10
creates noise to smoothe out the color banding. Doesn't look particularly good. Seems to introduce color artifacting at times.
- Floyd and Steingberg 5.50 MB Arbitrary Rating: 3/10
looks nearly identical to Heckbert. Might have slightly less artifacting, but still has some.
- Frankie Sierra dithering v2 4.41 MB Arbitrary Rating: 4/10
also creates noise, arguably more agressively, but the patterns seem to be more random, thus harder to notice, so it's less immediately obvious.
- Frankie Sierra dithering v2 "Lite" 6.11 MB Arbitrary Rating: 4/10
creates even more noise, but does it in a smarter way, where it looks like it takes turns color-banding and then noisy-dithering gradients. In some scenarios seems to be very good, mildly noisy most of the time, but can artifact pretty hard sometimes.
- Frankie Sierra dithering v3 3.82 MB Arbitrary Rating: 7/10
minimal noise, but also doesn't eliminate color-banding completely. Seems like a natural evolution of the "Lite" version, but now it's 90% color banding + 10% noisy-dithering.
- Burkes 4.49 MB Arbitrary Rating: 6/10
seems to apply Heckbert-style noise but the application style is closer to Sierra v3 - it's obviously noisy but only at very minor parts of the animation. Slightly better at smoothing out banding
- Atkinson 3.78 MB Arbitrary Rating: 8/10
very similar to Sierra v3, hard to tell them apart. Seems like an overall improvement - smoother and less obvious noise patterns.
- None 3.58 MB Arbitrary Rating: 10/10
just color-banding. No noise, outside of noise created by the compression. Honestly, no reason to go with anything else for cel shaded content
Spending the file size allowance on lanczos scaling to produce an infinitely better-looking picture than the bilinearly-smudged crap (even if you have to cut down the palette size) is a much better idea IMO.