Test setup was simple:
-- Reference encoders were Sony 8- and 10- bit raw YUV wrapped AVI, the most accurate we've tested over the years.
-- Reference source is a static Belle-Nuit chart.
-- Encodes in Handbrake were done at Baseline.
-- We were looking for mathematical, not necessarily perceptual losses, so PSNR is reasonably adequate to detect aggregate increases (or decreases) in Q-noise.
-- All PSNRs are referenced to the Source, not to each other.
-- The PSNR is 100% for 10-bit upconversions made in 8- and 32- bit float workspaces comparatively, proving no color-depth alchemy takes place; 0's are merely substituted for the last two bits, according to my partners.
1. Reference source is YUV 8 bit (100 MB)
2. mp4 from Handbrake PSNR 47.67, 54 KB (not bad)
3. convert to 10-bit YUV with no filtering PSNR 44.20 (oh it went down not up), 132 MB
4. next mp4 from Handbrake PSNR 43.62 (still not bad but not an improvement), 49 KB
This is enough to persuade me to truncate further testing and discussion. Call it an "early fail."
The tests were replicated from scratch to rule out settings errors, and the results were identical.
Conclusions:
1. 8 bit mp4 from Handbrake does not benefit measurably from upconverting to 10 bit YUV
2. Subsequent downconverting to 8 bit causes even more losses (I theorized this a couple of years back)
3. Net losses from the two-generation process were 4%.
2. Handbrake needs 10 bit encoding to be ready for the new formats, but NOT for the reasons being bandied about the internet.
3. Handbrake DOES benefit from being fed 10-bit native source over 8 bit source, and outperforms MainConcept in that arena, from the tests linked in the other thread. Upconverting native 8-bit to pseudo 10-bit source causes losses, not gains.
