4 ms·

I’ve recently compared WebP and AVIF with the reference encoders (and rav1e for lossy AVIF), and for similar quality, WebP is almost instant while AVIF takes mo

I’ve recently compared WebP and AVIF with the reference encoders (and rav1e for lossy AVIF), and for similar quality, WebP is almost instant while AVIF takes more than 20 seconds (1MP image). JXL is not yet widely supported, so I cannot really use it (videogame maps), but I hope its performance is similar to WebP with better quality, for the future.

adzm 9mo ago
You have to adjust the CPU used parameter, not just quality, for AVIF. Though it can indeed be slow it should not be that slow, especially for a 1mp image. The defaults usually use a higher CPU setting for some reason. I have modest infrastructure that generates 2MP AVIF in a hundred ms or so.
- LtdJorge 9mo ago
  I tested both WebP and AVIF with maximum CPU usage/effort. I have not tried the faster settings because I wanted the highest quality for small size, but for similar quality WebP blew AVIF out of the water. I also have both compiled with -O3 and -march=znver2 in GCC (same for rav1e's RUSTFLAGS) through my Gentoo profile.
  - adzm 9mo ago
    Maximum CPU between those two libs is not really comparable though. But quality is subjective and it sounds like webp worked best for you! Just saying though, there is little benefit in using the max CPU settings for avif. That's like comparing max CPU settings on zip vs xz!
fishgoesblub 9mo ago
rav1e has not had an actual update in performance or quality in years since funding got dropped. Use an encoder like aom, or svt-av1.
- LtdJorge 9mo ago
  I tried both AOM and rav1e, same quality with rav1e producing 20% larger image, same performance more or less (too long)
spartanatreyu 9mo ago
Other comments here are good, but one thing that's worth pointing out: Encoding time isn't as important as decoding time since encoding is generally a once-off operation. Yeah, we all want faster encodes, but the decodes are the most important part (especially in the web domain).
- LtdJorge 9mo ago
  I know, thats why I used max CPU settings. But when processing map tiles with a final total compressed size of half a terabyte, and each one is 200kB, taking 20s per tiles is prohibitively expensive.

adzm 9mo ago

You have to adjust the CPU used parameter, not just quality, for AVIF. Though it can indeed be slow it should not be that slow, especially for a 1mp image. The defaults usually use a higher CPU setting for some reason. I have modest infrastructure that generates 2MP AVIF in a hundred ms or so.

LtdJorge 9mo ago
I tested both WebP and AVIF with maximum CPU usage/effort. I have not tried the faster settings because I wanted the highest quality for small size, but for similar quality WebP blew AVIF out of the water. I also have both compiled with -O3 and -march=znver2 in GCC (same for rav1e's RUSTFLAGS) through my Gentoo profile.
- adzm 9mo ago
  Maximum CPU between those two libs is not really comparable though. But quality is subjective and it sounds like webp worked best for you! Just saying though, there is little benefit in using the max CPU settings for avif. That's like comparing max CPU settings on zip vs xz!

fishgoesblub 9mo ago

rav1e has not had an actual update in performance or quality in years since funding got dropped. Use an encoder like aom, or svt-av1.

LtdJorge 9mo ago
I tried both AOM and rav1e, same quality with rav1e producing 20% larger image, same performance more or less (too long)

spartanatreyu 9mo ago

Other comments here are good, but one thing that's worth pointing out: Encoding time isn't as important as decoding time since encoding is generally a once-off operation. Yeah, we all want faster encodes, but the decodes are the most important part (especially in the web domain).

LtdJorge 9mo ago
I know, thats why I used max CPU settings. But when processing map tiles with a final total compressed size of half a terabyte, and each one is 200kB, taking 20s per tiles is prohibitively expensive.