Why does x265 preset fast get the smallest file size by const. quality?

[Adsmt]

I tried many videos to test with x265. I set crf to 23.0 and found preset fast always get the smallest files than other presets.

Why do slow, slower or even very slow presets spend much more time than fast but still get bigger sizes?
Is constant quality real "constant" or it depends on different presets even though it is set to "the same" quality?

My x265 version is 2.0.0.2.

[jagabo]

The slower presets use more accurate motion estimation settings and deliver better quality. View the results and you'll see the difference. Look at the edges of moving objects.

[Adsmt]

So the same value of crf doesn't mean the same quality by different presets?

[jagabo]

No. Higher (slower) presets give better quality.

[-Habanero-]

jagabo, that wasn't what he was asking. But the answer is no, anyhow. CRF is a rough estimation of quality. CRF18 with slow settings won't give the same quality as CRF18 with fast settings. In one test I did, CRF20 with maximum settings was about the same quality as CRF26 with fast settings.

[abolibibelot]

In one test I did, CRF20 with maximum settings was about the same quality as CRF26 with fast settings.

It's late at night and I'm about to sleep, but shouldn't it be the opposite ? I.e., a higher CRF with slow settings producing a level of quality comparable with that of a lower CRF with fast settings. (I haven't tried x265 yet -- that'd be foolish on my 2009 computer -- and read very little information about it, but I guess CRF and quality are affecting each other in inverse direction just like with x264, while slow settings should improve quality as “jagabo” said.)

[-Habanero-]

It's late at night and I'm about to sleep, but shouldn't it be the opposite ? I.e., a higher CRF with slow settings producing a level of quality comparable with that of a lower CRF with fast settings. (I haven't tried x265 yet -- that'd be foolish on my 2009 computer -- and read very little information about it, but I guess CRF and quality are affecting each other in inverse direction just like with x264, while slow settings should improve quality as “jagabo” said.)

No, slow settings decrease the filesize because of the better compression they offer but the quality also suffers as a result. Pay attention now. The bitrate is greatly reduced and the quality only marginally reduced. So to compensate, a lower CRF must be given. The CRF26 with fast settings had a much higher bitrate than the CRF18 with max settings yet the quality was the same.

The same CRF is supposed to give the same quality regardless of the configuration but in practice it doesn't work out that way.

[jagabo]

No, slow settings decrease the filesize because of the better compression they offer but the quality also suffers as a result.

No. If you take a particular video and run CRF encodings at each preset you'll find that file sizes decrease dramatically (~50 percent) from utrafast to veryfast. Then as you go to higher presets they rise again by a small amount (~10 percent). Until about medium where they start to fall again by a small amount. And if you look at the visual quality you'll see that it increases from veryfast to placebo.

For example, I just encoded a short video at CRF 18 with x264:

ultrafast: 32.8 MB
superfast: 22.5 MB
veryfast:  15.7 MB
faster:    17.3 MB
fast:      17.5 MB
medium:    16.9 MB
slow:      16.6 MB
slower:    16.5 MB
veryslow:  15.3 MB 
placebo:   15.7 MB

Of course, the exact size variation and inflection points vary from video to video but the general trend holds. I haven't used x265 much but I believe it behaves similarly.

But abolibibelot, don't take my word for this. Try it yourself.

[-Habanero-]

ultrafast: 32.8 MB
superfast: 22.5 MB
veryfast: 15.7 MB
faster: 17.3 MB
fast: 17.5 MB
medium: 16.9 MB
slow: 16.6 MB
slower: 16.5 MB
veryslow: 15.3 MB
placebo: 15.7 MB

..

[jagabo]

Drawing a trend line doesn't preclude the fact that there is a dip at veryfast and it takes several more presets before the size dips down below that. Nobody's arguing that the general trend isn't downward.

[sophisticles]

No. Higher (slower) presets give better quality.

This is exactly opposite of what you, and many others, have said about the relationship between CRF and presets in the past.

The mantra has always been that CRF denotes a given quality and the presets determine the final size; i.e. CRF 18 + ultra fast will produce the same quality as CRF 18 + placebo only the former will use more bit rate to accomplish the same quality.

In this thread, you, and others, have completely contradicted a decade of accepted CRF theory. If as you say at slower presets CRF ups the bit rate then CRF is, as I have always believed, pretty much useless.

Here's the explanation for what the OP is observing, the slower presets make use of various psycho-visual enhancements that faster presets do not and as you get to the slower settings the psycho-visual settings get more aggressive as well. If you read up on x265's psycho-visual algorithms, and x264's for that matter, you find that various settings will actually pump up the bit rate in CRF mode in order to avoid artifacts.

On the x265 website there is an explanation related to psy-rd and one other psy setting whose name escapes me at the moment, where it states that if you set the associated settings too high you get artifacts due to portions of the picture being bit rate starved as the algorithm tried to move bits around. To prevent this, in CRF mode, setting the psy optimizations too high will balloon up the bit rate to try and avoid the artifacts.

The takeaway here is that x265 is a steaming pile of garbage that shouldn't be used by anyone for anything, unless you have a raw sewage fetish.

Thank you.

[sneaker]

The mantra has always been that CRF denotes a given quality and the presets determine the final size; i.e. CRF 18 + ultra fast will produce the same quality as CRF 18 + placebo only the former will use more bit rate to accomplish the same quality.

In this thread, you, and others, have completely contradicted a decade of accepted CRF theory. If as you say at slower presets CRF ups the bit rate then CRF is, as I have always believed, pretty much useless.

It wasn't accepted theory, only a common misconception. CRF is not an absolute measurement of quality. It wasn't that way with x264 either. It doesn't make it useless, it just means you shouldn't use CRF to compare different settings without taking the bitrate into the equation.

[SameSelf]

The takeaway here is that x265 is a steaming pile of garbage that shouldn't be used by anyone for anything, unless you have a raw sewage fetish.

Wow. Didn't see that one coming. Should I be afraid of HEVC?

[-Habanero-]

In this thread, you, and others, have completely contradicted a decade of accepted CRF theory. If as you say at slower presets CRF ups the bit rate then CRF is, as I have always believed, pretty much useless.

deadrats, I used to believe CRF was a great indicator of quality too until I did a test and proved myself wrong. It's irritating that in the software world of 0s and 1s when something either works or doesn't STILL has whackjobs like you who find a way to turn it into a f*cking abstract, religious debate where everything is unfalsifiable.

Also, since I don't use psy, that is irrelevant. Psy INCREASES artifacts at lower bitrates anyway.

[sophisticles]

The takeaway here is that x265 is a steaming pile of garbage that shouldn't be used by anyone for anything, unless you have a raw sewage fetish.

Wow. Didn't see that one coming. Should I be afraid of HEVC?

It was late and i had just worked a 14 hour day when I wrote that; I would like to take this opportunity to apologies to steaming piles of garbage I may have offended and those within the raw sewage fetish community, most are fine, decent folks and they do not deserve to have their good names sullied by being associated with x265.

[sophisticles]

In this thread, you, and others, have completely contradicted a decade of accepted CRF theory. If as you say at slower presets CRF ups the bit rate then CRF is, as I have always believed, pretty much useless.

deadrats, I used to believe CRF was a great indicator of quality too until I did a test and proved myself wrong. It's irritating that in the software world of 0s and 1s when something either works or doesn't STILL has whackjobs like you who find a way to turn it into a f*cking abstract, religious debate where everything is unfalsifiable.

Also, since I don't use psy, that is irrelevant. Psy INCREASES artifacts at lower bitrates anyway.

My apologies, but what does rodents that are no longer alive have to do with this discussion? Is it because one would expect to find substantial numbers of them in steaming piles of garbage?

[KarMa]

If as you say at slower presets CRF ups the bit rate then CRF is, as I have always believed, pretty much useless.

I think jagabo was agreeing that bitrate drops on average, the slower the preset for a given CRF. Not the other way around which you seem to have misread.

[jagabo]

If as you say at slower presets CRF ups the bit rate then CRF is, as I have always believed, pretty much useless.

I think jagabo was agreeing that bitrate drops on average, the slower the preset for a given CRF. Not the other way around which you seem to have misread.

It's just that the veryfast preset (x264) is anomalous in this respect.

And obviously, "CRF=contant quality" is a simplification. The image quality increases a bit with slower presets, but the difference isn't huge. And it's certainly closer to constant quality than encoding every video at the same arbitrary bitrate -- the context in which this is usually discussed.

[knightplex]

I just did a CRF Encoding at 26. My speed was medium. When I decreased preset to "slow", the bitrate and the file size both increased. Actually only bitrate should have increase and file size should have decreased. But that didn't happen. I further changed the preset to "slower" the file size and bitrate both increased again. Why is this happening? I used Staxrip latest version 10 bit x265.

[jagabo]

Actually only bitrate should have increase and file size should have decreased.

File size and bitrate are simply two ways of measuring the same thing.

file size = bitrate * running time

If the bitrate goes up so does the file size.

[sneaker]

...

[knightplex]

I know, but I was under the impression that Preset was used to only influence compression. Slower the preset, better the compression. That's not happening at all.
And that is what the software says too.

[sneaker]

"Better compression" doesn't mean "lower filesize". It means "lower filesize at the same quality". Since same CRF does not guarantee same quality you can't conclude from your test this principle has been violated.

[knightplex]

Same CRF value doesn't guarantee same quality?

[jagabo]

Same CRF value doesn't guarantee same quality?

How many times does this have to be said? Different presets at the same CRF do not deliver exactly the same quality. All you have to do is look at a video compressed at veryfast and a video compressed at veryslow and it will be obvious. The latter will be more true to the source. It will have smoother edges on moving objects and better retention of small, low contrast details.

[knightplex]

Okay, one more thing. Currently I used CRF 22 with medium speed.
Now which quality will be better, CRF 22 with slower preset or CRF 20 with medium preset. I mean just your guess on how it will affect if I just lower CRF value, or keep the same CRF value and make the speed lower. Which would be a better choice?

[sneaker]

Most of the time lower CRF will look better. But it's not guaranteed.

[jagabo]

Okay, one more thing. Currently I used CRF 22 with medium speed.
Now which quality will be better, CRF 22 with slower preset or CRF 20 with medium preset. I mean just your guess on how it will affect if I just lower CRF value, or keep the same CRF value and make the speed lower. Which would be a better choice?

I agree with sneaker. Most of the time CRF 20 at medium will look better. And the file will usually be significantly bigger. Though "better" is somewhat subjective. Different people may weigh different flaws differently.

But why take anybody's word for it? Just run the test yourself.

[Detmek]

I did small test with x265-8bit, CRF 22 to measure differences in bitrate, quality (SSIM) and encoding time. I used Simple x264/x265 launcher with command: --preset xxx --tune ssim --ssim. I replaced x265 build with 2.0.52. Test was done on 720p clip.

x265 [info]: HEVC encoder version 2.0+52-6ee705ecb433
x265 [info]: build info [Windows][GCC 5.3.1][64 bit] 8bit
x265 [info]: using cpu capabilities: MMX2 SSE2Fast SSSE3 SSE4.2 LZCNT

P.S. My CPU (Pentium G3220) does not support AVX/AVX2.



It seams there is a bug in preset optimization with veryfast and faster because encoding time and quality is almost equal. Same thing happend with 2.0.2 build that comes with Simple x264/x265 launcher.

[Epaminaidos]

Just for completeness sakes: The settings leave me as puzzled as you all.
I just completed a test myself and found this thread shortly after.
My setup:
Core-i7-6700HQ, 8GB RAM
Source: 30s of 1080p60 encoded in H.264@150Mbps
Tool: Handbrake 0.10.5 (includes x265 v1.9)
Settings: RF12

Results (I did not wait for "slower" and the slower ones to finish):


I don't know, what to make from that