x264 --direct

[simps]

What would give better quality?

--direct auto
--direct spatial
--direct temporal

I read on doom9 x264 guide, that people would believe temporal to be better quality. Is it? What about the auto option? Shouldn't it be better?

[]'s
Simps

[Dark Shikari]

Auto is best.

Temporal can sometimes give better visual quality than spatial, but it is broken with b-pyramid (auto will correctly select which one is better, temporal will force it where it makes no sense).

[burfadel]

Yeah, based on really old suggestions people still think forcing spatial (or temporal) is better with 'typical' settings (such as I would classify b-pyramid as).

[Forteen88]

Temporal can sometimes give better visual quality than spatial, but it is broken with b-pyramidWhat do you mean by 'broken'?
EDIT: Thanks for the explanation DS.

[Dark Shikari]

What do you mean by 'broken'?Due to a flaw in the H.264 specification, temporal direct mode becomes meaningless in some of the B-frames in complex structures such as B-pyramids.

In particular:

P
b <-Doesn't work
B <-Works
b <-Doesn't work
P

In 2-pass, auto will take care of this. In 1-pass, it just won't use temporal in most cases.

[simps]

True. On my 1st pass using --direct auto, I believe x264 said it used 2% temporal only I think. I'm doing 2nd pass now, will report how it goes.

[burfadel]

Would a frame type decision lookahead help resolve that?

[Dark Shikari]

Would a frame type decision lookahead help resolve that?In theory one could shoehorn this into the lookahead, but I'm not sure the lowres lookahead would make a particularly accurate decision.

[Forteen88]

... In 2-pass, auto will take care of this. In 1-pass, it just won't use temporal in most cases.What about CRF, does it have this problem as well? Should I rather use CRF 2-pass for better --direct auto?

EDIT: Ok, thanks DS, I thought you meant only ABR 1-pass.

[Dark Shikari]

What about CRF, does it have this problem as well? Should I rather use CRF 2-pass for better --direct auto?CRF is 1-pass. --direct auto helps very little overall, so it's really not worth bothering with going for 2-pass solely for that.

[burfadel]

If a lookahead were implemented for 'frame type decision', couldn't the results be reused, instead of a re-analysis?

[Dark Shikari]

If a lookahead were implemented for 'frame type decision'"were" implemented? This isn't 2004, we've had frametype lookahead for almost half a decade now.

[burfadel]

Ah yes sorry! I did read the earlier messages :) I meant if a suitably high quality lookahead were implemented, as to your earlier statement that it was low-res.

[Sagekilla]

It's a low res lookahead for a reason: That makes it -much- faster since you have less pixels to analyze.

[burfadel]

I realise that :) I guess I might be thinking something that can't be done, such that if a high res lookahead is used the results of the analysis can be used for the encoding process, instead of a re-analysis occurring.

[Sagekilla]

Another issue with having higher res lookahead is you'd be exacerbating the single threaded nature of it on multi-core systems. It might be neat to have a higher res option if we had multithreaded lookahead, but that might not be for a while :(