The Holy Grail settings of NVENC HEVC.Can you find better preset? (CONTEST)

[Abdler]

Hello!

The Holy Grail Settings of NVENC HEVC, Can you find better general preset for low bitrate (2500 Kbit/s) video?

If you found, please share it!

I used the latest Hybrid , Rigaya's NVENC with the following presets:

NVEnc --avhw -i -INPUT- --fps 23.976 --codec h265 --profile main10 --level 6.2 --tier high --sar 1:1 --lookahead 32 --output-depth 10 --vbrhq 2500 --max-bitrate 240000 --gop-len 600 --strict-gop --ref 7 --multiref-l0 7 --multiref-l1 7 --nonrefp --bframes 1 --bref-mode disabled --no-b-adapt --mv-precision Q-pel --cu-max 32 --cu-min 8 --preset quality --colormatrix bt709 --vpp-resize default --output-res 1920x1080 --vpp-gauss disabled --cuda-schedule auto --keyfile GENERATED_KEY_FILE --psnr --ssim --output "C:\Users\007\AppData\Local\Temp\2 b solid.265"

I also used Rigaya's built-in PSNR and SSIM measurement tool too.

[Abdler]

Any idea?

[Selur]

--bframes 1 --bref-mode disabled

have you tried enabling bref-mode and using more b-frames?

[Abdler]

--bframes 1 --bref-mode disabled

have you tried enabling bref-mode and using more b-frames?

I made 1b frame 2 b frame and 3b frame encodings, with two variations: 1 / with adaptive B-frame 2/ with fixed (solid) b frames.

I uploaded here the results:
https://sendgb.com/M0YSreEQlWX

Here is the original source file: https://sendgb.com/8pCny1KD76d

Is the fixed (solid) 1B frames the best version?

[Abdler]

SOLID (FIXED) 1 B frame results:
SSIM YUV: 0.959519 (13.927510), 0.987063 (18.881730), 0.987414 (19.001062), All: 0.968759 (15.052747), (Frames: 4922)
PSNR YUV: 41.852698, 47.663324, 47.616219, Avg: 43.075502, (Frames: 4922)
__________________________________________________ __________________

Adaptive 1 B frame results:
SSIM YUV: 0.959268 (13.900629), 0.987143 (18.908510), 0.987484 (19.025379), All: 0.968616 (15.032970), (Frames: 4922)
PSNR YUV: 41.782242, 47.695861, 47.645941, Avg: 43.016762, (Frames: 4922)
__________________________________________________ _______________________
SOLID 2 B frame results
SSIM YUV: 0.959184 (13.891746), 0.987477 (19.022821), 0.987908 (19.174925), All: 0.968687 (15.042762), (Frames: 4922)
PSNR YUV: 41.775293, 47.845018, 47.824288, Avg: 43.028969, (Frames: 4922)
__________________________________________________ _________________________________
Adaptive 2B frame results:
SSIM YUV: 0.959268 (13.900629), 0.987143 (18.908510), 0.987484 (19.025379), All: 0.968616 (15.032970), (Frames: 4922)
PSNR YUV: 41.782242, 47.695861, 47.645941, Avg: 43.016762, (Frames: 4922)
__________________________________________________ _____________________________________
Solid 3B frame results:
SSIM YUV: 0.958440 (13.813271), 0.987824 (19.144857), 0.988285 (19.312587), All: 0.968312 (14.991001), (Frames: 4922)
PSNR YUV: 41.629757, 47.991645, 47.985072, Avg: 42.915426, (Frames: 4922)
__________________________________________________ _____________________________________
Adaptive 3 b frame results:
SSIM YUV: 0.958726 (13.843203), 0.987423 (19.004305), 0.987791 (19.133101), All: 0.968353 (14.996646), (Frames: 4922)
PSNR YUV: 41.657191, 47.811351, 47.769798, Avg: 42.918961, (Frames: 4922)

[Zero-11]

make it preset slow

[Abdler]

make it preset slow

Is it possible in Rigaya's version of NVENC for the "preset slow" ?

Here is Rigaya's NVENC option list: https://github.com/rigaya/NVEnc/blob/master/NVEncC_Options.en.md

[Selur]

Have you actually looked at the streams? Do not trust the metric, trust your eyes.

[Abdler]

Have you actually looked at the streams? Do not trust the metric, trust your eyes.

Hello Selur!

Which video is the best according to your visual test?

[Selur]

doesn't matter, quality is in the eye of the beholder

[Abdler]

doesn't matter, quality is in the eye of the beholder

Yes, but you have also eyes, thus you have also an opinion...

[Zero-11]

make it preset slow

Is it possible in Rigaya's version of NVENC for the "preset slow" ?

Here is Rigaya's NVENC option list: https://github.com/rigaya/NVEnc/blob/master/NVEncC_Options.en.md

it says you are using b-frames which indicates a Turing card, fast/medium/slow are the new Turing settings which have better quality
if Rigaya's version doesn't support it try Handbrake

[Abdler]

make it preset slow

Is it possible in Rigaya's version of NVENC for the "preset slow" ?

Here is Rigaya's NVENC option list: https://github.com/rigaya/NVEnc/blob/master/NVEncC_Options.en.md

it says you are using b-frames which indicates a Turing card, fast/medium/slow are the new Turing settings which have better quality
if Rigaya's version doesn't support it try Handbrake

But it uses FFMPEG, which is infamous from its bad quality and slow encoding speed among the HW encoders. FFMPEG has not even proper HW decoder for input videos. In FFMPEG, everything goes trough the system memory and CPU, which makes the process backward/outdated.

[Selur]

Handbrake, FFmpeg and NVEncC usually all use the same libraries (from the current NVIDIA SDK) to access the VPU on your graphic card, so unless someone messes up, the quality should not differ.
It's just a difference how NVEncC and Handbrake name an option (and if they offer it to the user).
In case of the presets, Handbrake offers presets named 'fast/medium/slow' while NVEncC names them "performance/default/quality".

Cu Selur

[Zero-11]

In case of the presets, Handbrake offers presets named 'fast/medium/slow' while NVEncC names them "performance/default/quality".

Cu Selur

yea right. that's why the 1st ones have better quality, Handbrake didn't name them different. it has all of them

[Abdler]

In case of the presets, Handbrake offers presets named 'fast/medium/slow' while NVEncC names them "performance/default/quality".

Cu Selur

yea right. that's why the 1st ones have better quality, Handbrake didn't name them different. it has all of them

But Handbrake uses only FFMPEG, and FFMPEG has very few options for HW encoders.

[Selur]

Handbrake does not use FFmpeg.
It uses libav and other libraries and since libav is originated in FFmpeg sometimes folks write Handbrake uses FFmpeg, which it does not.
Like I wrote before both FFmpeg, Handbrake and NVEncC all use the same libaries, it's just a question of which option they support and how.

Cu Selur