about the performance of x265_2.8 compared to x265_2.0

[Former user]

Hi, I recently test the performance of x265 at version 2.8. I use x265_2.0 as baseline, and try to calculate bd-psnr & bd-ssim. I use CQP ranged from 25~40, step 5, and test four preset which are medium, slow, slower, veryslow.

I just test six video at 1920x1080 size. they name are below: 1. BasketballDrive_1920x1080_50 2. BQTerrace_1920x1080_60 3. Cactus_1920x1080_50 4. Kimono1_1920x1080_24 5. ParkScene_1920x1080_24 6. Tennis_1920x1080_24

the result is amazing me, I find both bd-psnr & bd-ssim of version 2.8 is worse than 2.0. I post my test result in the attachment. can anybody help me to find out the reason?

[Pradeep Ramachandran]

Thanks for the report. Your result does seem strange to us. Please give us some time to look into it and get back.

[Jung Zhao]

Hi, I forget to paste my options. below is my test parameters for both 2.8 and 2.0:

--preset=medium|slow|slower|veryslow --qp=25|30|35|40 --fps

and each source I just encode 200 frames.

[Nik McNulty]

This is almost certainly due to both the philosophical and developmental focus on SUBJECTIVE visual quality over OBJECTIVE metric scores.

The new (but maybe not improved) lambda tables in v2.4 are a particular example of this.

To revert to a quantitative schema would require abandoning x265's entire concept of "perceived" quality through psycho-visual "optimisations".

[Jung Zhao]

Thank you for repling Nik McNulty.

According to your writing, is that mean x265 team caring about SUBJECTIVE visual quality more than OBJECTIVE socre when they publish new version of x265.

I am very curious at how do the team to assess the SUBJECTIVE visual quality? use VMAF or eyes?

[AlexSuper]

You should read the x265.org about How to compare video encoders The --tune psnr and --tune ssim are better for non-psychovisual metrics, try try see.