Video encoding with Nvidia HEVC.

[Fox_25]

Hello,
I got a cheap GTX 1660 super and I've „rushed“ with FFmpeg batch to reduce the size of many many videos, preferably without losing quality of course....


-c:v hevc_nvenc -rc constqp -qp 28 -preset p7 -tune hq -pix_fmt yuv420p -rc-lookahead 4 -bf 1 -b_ref_mode 1 -keyint_min 1 -multipass 2 -spatial_aq 1 -aq-strength 10 -temporal_aq 1 -refs 7 -qdiff 20 -qcomp 0.9 -me_method umh -subq 9 -me_range 36 -b:v 0 -g 300


I found one x264 „preset“ that I modified for HEVC and I'm just tweaking the qp for optimal quality and file size. I wonder if there's anything else to improve on it and if it's ok, but according to the NVIDIA NVENC hevc encoder „datasheet“ probably not, at most I could try how constant bitrate will behave. However, I don't really want to do that because the results are good in my opinion and the settings may not be ideal for constant bitrate.

Orig, avg bitrate 3.6MB/s, File size 573MB
https://abload.de/img/or37fc6.png

avg. bitrate 1.7MB/s, File size 270MB, qp 27, spatial aq strength 8
https://abload.de/img/aqtpgxcmz.png

avg. bitrate 860kB/s, File size 147MB, qp 30, spatial aq strength 6
https://abload.de/img/qp30z4fkn.png


Does that look okay to you? So far I have the best results with qp28 and Spatial AQ 10. Qp 27 and qp28 with Spatial AQ 10 have identical screen quality, or more specifically, I couldn't find the difference in the picture, only bigger file.

avg. bitrate 1.57MB/s, File size 242MB, qp 28, spatial aq strength 10
https://abload.de/img/qp28cmdb0.png



Thanks for response.

[lollo]

to reduce the size of many many videos, preferably without losing quality of course

If that would be possible, then the original size of the videos could be what you aim to, of course

[Fox_25]

Of course, this was meant with exaggeration because any decoding>encoding has a quality loss, especially if I want to reduce the file size. My point was if my results are good or if there is something else to improve.

[lollo]

My point was if my results are good or if there is something else to improve.

The first, but this is based just on fixed sample images