rd-refine will cause some glitch in some encoded frames.

[Z]

The combination is:

--max-tu-size <32

--vbv-maxrate and --vbv-bufsize with arbitrary value (so it not need to be effective, just need to be there?)

--rd-refine (with --rd 5 of course, but --rd 5 alone is fine.)

--limit-refs <2

--weightb

Other parameters may make it worse or better, but it is believed that the combination above is the key leading to this bug. (the strikethrough parameters are just fell into this category)

Test source and two example encoding command line is included. The glitch will mostly appear in frame 427~431.

Images contain results from different “other parameters“.

Based on more testing and discussion in links mentioned in comments, rd-refine is the only one real key to this bug.

‌

‌

‌

[Z]

Will this issue end up like another one that nobody cares for over two years?

[khaf]

I have tried the combination in his pack, the problem happened as he said.

The Screenshot: https://slow.pics/c/eKfO0A9x#

[khaf]

And I appeared that many people used the combination were similar to that, the similar(or same?) problem has happened, especially in fades and high dynamic scenes.

Such as:

https://bbs.acgrip.com/thread-10120-1-1.html

https://bbs.acgrip.com/thread-10159-1-1.html

‌

[Z]

I tried to do something and there’s more testing needed, if it works I’ll create a PR.

As for speed and quality,my testing shows it makes encoding a little slower, but improves quality rather than the opposite. (this opposite effect may be just a weird result in this single testing)

Because traditional metrics are not so good at revealing the glitch reported, manual examinations are required.

https://bitbucket.org/mister-z/x265_git/commits/8d78b0fa48d09f0f2b5a58372d573a4711a432fa

‌

[Z]

A few more speed&quality results, I’ll just post urls

https://i.imgur.com/SmGjCjO.png

https://i.imgur.com/wXkDFOx.png

https://i.imgur.com/gY4RmF9.png

‌