Recompressing jpeg lossless

[kaitz]

Is there difference recompressing jpeg when source is artiffical vs photo image?

I tested with paq8pxd modified jpeg model and gain is better to get with artifical jpeg files then photo jpeg.
Same is with artifical bitmap images, but why?

Almost no speed penalty but 2k smaller file on 3m artifical jpeg image. No gain on photo.

[Shelwien]

Artificial images contain less noise and have simpler structure in general, easier to predict.
Significantly downscaling a CG image (like 4x or more) might make it "equal" to a photo.
Its also rather common to ignore artificial images in image codec benchmarks, because they behave differently.
With high enough jpeg quality (and with iDCT contexts in paq jpeg model) its reasonable to expect same differences in jpeg recompression.

[Gonzalo]

JPGs can be quite different depending on the encoder. Try optimizing them with ect then compare again and see if the difference is the same

[Jyrki Alakuijala]

JPGs can be quite different depending on the encoder. Try optimizing them with ect then compare again and see if the difference is the same

On Brunsli my experience is that normal photographs compressed initially with libjpeg compress further 22 %.

When I have done the lossy phase with Guetzli instead of libjpeg (with a 30+ % win in highest qualities), I only get 14 % of additional compression from Brunsli.