x265 HEVC Encoder

[LigH]

Are you still speculating why x265 with broken psychovisual enhancements looks worse? I don't remember that they were announced as "fixed".

Some more material to watch (http://www.mediafire.com/?6lfp2jlygogwa): "Johnny" 720p with preset slower and psy-rd 0.0 .. 1.0 in steps of 0.2. I have two more similar clips, any suggestions which combinations of options I should try?

[Atak_Snajpera]

Certainely that xvid will provide excellente quality too here.
Latest Xvid + slowest settings enabled + 2-pass
https://mega.co.nz/#!IZFmBDqI!4oIPL_XKkvac0-X6L-FRUsOz_z97RR025T7e1W_ZWAA

frame 6310
Xvid -> http://i.imgur.com/B7SXQOa.png

frame 7380
Xvid -> http://i.imgur.com/lIfJWMG.png

frame 12630
Xvid -> http://i.imgur.com/9vCdHCs.png


Indeed "excellente quality"

http://i.imgur.com/sVD4nNH.png
http://i.imgur.com/RdqE8fr.png
http://i.imgur.com/qxQPSRG.png

x264 will turn into a block madness, x265 will look like HD vs SD when you compare the two.
Both will look ugly. x264 will give you blocks while x265 will "apply" ultra strong denoising filter. In both cases important details will be sent to hell ;)

Similar history with audio compression. 48Khz HE-AAC@32kbps with parametric stereo will definitely sound better than mp3 with the same bitrate and so what. The question is : Will you be encoding your FLAC collection for your ipod/smatphone using such ridiculously low bitrate??? I do not think so! Most likely you will be using something between 64kbps-128kbps.

[LigH]

Both will look ugly. x264 will give you blocks while x265 will "apply" ultra strong denoising filter. In both cases important details will be sent to hell ;)

Already in April I uploaded files to compare: "in_to_tree" 1080p50, encoded with x265 --crf 30 (http://www.mediafire.com/watch/g80s1nszbjqp8po/in_to_tree_1080p50.x265-crf30.mp4) and with x264 in 2-pass using the same (http://www.mediafire.com/watch/65zm6z6ct6u3jgk/in_to_tree_1080p50.x264_br~x265-crf30.mp4) / 2 times (http://www.mediafire.com/watch/l2a9efbuzbvudbj/in_to_tree_1080p50.x264_br~x265-crf30_x2.mp4) / 3 times (http://www.mediafire.com/watch/pshba3nuef6baut/in_to_tree_1080p50.x264_br~x265-crf30_x3.mp4) the bitrate of the x265 encode (each with preset "veryslow"). And I believe I should try again soon.

[Atak_Snajpera]

@LigH
Try to encode full 2h movie 1920x1080 (16:9) with x265's preset very slow ;) I do not think that end user will leave his PC on for whole week to encode just one movie ;) Do you see where I'm going? Veryslow preset is insanely sloooow to be practically useful. Not to mention about electricity bills ;) Magic tricks won't speed up encoding to sane levels.

[nevcairiel]

They could change the presets to be faster at the cost of quality, or you could just use a faster preset all by yourself..
You shouldnt be surprised if the "veryslow" preset is ... very slow!

There was never any argument that x265 would be as fast as x264, everyone knows that its going to be slower.

[Atak_Snajpera]

Test sample : in_to_tree_1080p50.y4m

x265 --crf 30 --preset veryslow --psy-rd 1.0 (average bitrate 3279 kbps) - ENCODING TIME : 6m58s
https://mega.co.nz/#!tV8QEQIC!KXIHtFHO5ABKKl9rPuefZA_DzUuobA8dH5uQ8bq2G88

x264 --crf 29 --preset veryslow (average bitrate 3351 kbps) - ENCODING TIME : 0m49s
https://mega.co.nz/#!lY0jkTwC!y12ljhXV6-roO9aLkJVjVd-GPbLidcyq-zNsjYoAFCY

Frame 50
uncompressed -> http://i.cubeupload.com/FMmLDC.png
x265 -> http://i.cubeupload.com/0cxHBh.png
x264 -> http://i.cubeupload.com/DPAfrZ.png

This is perfect example where details are gone. No matter what encoder you will use.

[LigH]

There was probably a similar trade-off discussion between H.264 and VC-1, I just don't remember in detail anymore, but I believe that WMV used more elaborate post-processing filters to cover compression artefacts.

Next time I will probably compare a faster x265 preset against the veryslow x264 preset; but I am not sure when and which exactly, it may depend on the development of psy-rd preset limits in x265, and a prerequisite will be that psy-rd will work correctly. Which is not yet the case, as far as I remember. When it is fixed, it may even work with smaller rd levels; surrendering psychovisual features because they are too slow would limit practical use of x265 a lot. Let's be curious and look forward...

[LigH]

Even more material to watch (http://www.mediafire.com/?6lfp2jlygogwa): "KristenAndSara" 720p with preset veryslow and psy-rd 0.0 .. 1.0 in steps of 0.2.

The artefacts above the left shoulder (Frame ~350) appear to be mainly a temporal development.

http://frupic.frubar.net/shots/31892.jpg

[Audionut]

@LigH
Magic tricks won't speed up encoding to sane levels.

Some people call it development (http://forum.doom9.org/showthread.php?t=141352)!

What reasonable expectation do you put on x265 becoming "mainstream"? 2 years? There's another 20% or so speed with hardware alone.

I really wish you would stop expecting x265 to be a polished product, here and now. Allow the devs some time to optimise it.

[fumoffu]

Good news everyone!
I accidentally discovered how do x265 encodes much faster! Basically you take x264 and set aq-strength to some low value like 0.3, you can also turn down psy-rd to similar value. Now you get files that are much smaller (if using CRF) and have very nice edges but unfortunately loose some detail... I was also using preset slow, brames 5 and 0 trellis but that is not too important.
I'm half jocking here but I really was surprised how similar the results are! I suspect those settings could help x264 compete with x265 at those very low bitrates everybody is excited about, come on - give it a try :)

[Atak_Snajpera]

@Audionut
Honeymoon is over! Time to check what we have now. I'm not going to sit and wait for your permission if i want to do some tests. Get used to it.

[x265_Project]

I really wish you would stop expecting x265 to be a polished product, here and now. Allow the devs some time to optimise it.
I don't care how, I want it NOW!!! (https://www.youtube.com/watch?v=TRTkCHE1sS4&feature=youtu.be&t=1m54s)

[LigH]

Especially a "sniper" :sly: should know that patience is a virtue... :p

[x265_Project]

But seriously folks. We appreciate the passion and the interest. I can assure you that we're working very hard every day to continue to improve x265 in three ways;

Encoding efficiency (subjective visual quality / bit rate)
Performance
Features (rate control, etc.)

Encoding efficiency is our highest priority. It doesn't matter if it goes fast if it doesn't produce the best possible visual quality at any desired bit rate. Our development team is still working to optimize psy-rd, and they are exploring a number of possible improvements to the core algorithms that x265 uses. We'll make these improvements available as soon as they are tested and validated.

Performance is a high priority, and we have a very solid roadmap focused on performance improvements. I don't want to share the details at this point, as we're in discussions with many companies regarding the funding and the priority for these improvements. There are many possible ways to improve performance. Generally, we're working to improve performance without sacrificing efficiency or capabilities.

Features like rate control are always important, especially to our commercial customers. We know where the opportunities for improvement lie, and we have developers who are working on these improvements.

[Sagittaire]

Latest Xvid + slowest settings enabled + 2-pass
https://mega.co.nz/#!IZFmBDqI!4oIPL_XKkvac0-X6L-FRUsOz_z97RR025T7e1W_ZWAA

frame 6310
Xvid -> http://i.imgur.com/B7SXQOa.png

frame 7380
Xvid -> http://i.imgur.com/lIfJWMG.png

frame 12630
Xvid -> http://i.imgur.com/9vCdHCs.png


Indeed "excellente quality"

http://i.imgur.com/sVD4nNH.png
http://i.imgur.com/RdqE8fr.png
http://i.imgur.com/qxQPSRG.png


Your conclusions are simply ridiculous. I see your sample with xvid codec. And quality is simply, for my eyes, something like DVD quality. For me it's high quality ... at this bitrate. Why use x265 or even x264 here if xvid make really good job?

Your prolem is that you want absolutely prove with really strange test that x264 is superior at x265.

I make many test too. And the reality is that x265 is really superior, and by far, at x264 in difficult situation, in most case. At low quantizer, only expert eyes like your or mine, will see real advantage for x264. For high quantizer, x265 will produce more pleasure result than x264 for really large majority of people.

Make the same encoding at 500 Kbps or even 250 Kbps and see the result. It's not necessary to upload your sample because I know the result.

Please, Atak_Snajpera, now, stop your trooling ...

[foxyshadis]

Atak_Snajpera, I think you should start a new thread for ongoing comparisons to other codecs (and it'd be helpful to include other codecs than x264 too), or I'll have to split it. This thread should be more about announcements, help using it, and discussions about the ongoing development. The constant comparisons and arguing are really derailing this thread.

My own two cents are that I remember when x264 took days to encode a full length movie, and that was fast when I'd just stepped up from a Pentium M to a Core Duo; now x265 takes the same on my Haswell i7. That's the nature of progress, and why H.264 took so many years to catch on. VP9 is likewise slow as hell right now too, they will slowly become more mainstream as hardware catches up.

[nhakobian]

I don't care how, I want it NOW!!! (https://www.youtube.com/watch?v=TRTkCHE1sS4&feature=youtu.be&t=1m54s)

I think this one might also be slightly relevant :D:
https://www.youtube.com/watch?feature=player_detailpage&v=Yrst8gmd9Cw#t=24

Back on topic:

I think its impressive how much work you guys have done in such little time. Building on the framework of x264 must have saved some time, but its good seeing someone actively working on new open source encoders. Best of luck!

[Guest]

@Sagittaire

Please don't trigger flamewars by calling someone a troll. Atak_Snajpera is a highly valued forum member who has made significant contributions.

Let's all keep it technical, please. Thank you.

[foxyshadis]

Seriously, guys, get over it. (OK, personal attack posts are removed now.)

As far as I'm concerned, any comparison posts are basically useless unless you have all of: Source file (and/or script), encoder version and 8/10 bit, encoder command line, and output sequences. Unless you're showing artifacts, stop posting full frames, they're arbitrary at best and totally misleading at worst, and you need a minimum of a few in a row from a few separate scenes to really tell a story with any confidence. Let the output speak for itself, and let's keep the emotion and ego out of it.

Moving on: I figured out the error in the KristenAndSara sequence; it's so noisy but quantized so flat that the psy-rd is thrown off and tries to pull that noise energy from wherever it can, in this case the hair. psy-rd 1.0 is way too much; it's actually trying to match MORE noise energy than exists in the original area, and totally mispredicting. It's much less ugly with 0.2 or 0.4, although it's apparent that psy-rd is also just a bad idea with very low bitrates. (Makes sense.) The effect completely disappears with --nr 100 even at --psy-rd 1.0.

[x265_Project]

As far as I'm concerned, any comparison posts are basically useless unless you have all of: Source file (and/or script), encoder version and 8/10 bit, encoder command line, and output sequences. Unless you're showing artifacts, stop posting full frames, they're arbitrary at best and totally misleading at worst, and you need a minimum of a few in a row from a few separate scenes to really tell a story with any confidence. Let the output speak for itself, and let's keep the emotion and ego out of it.

Agree completely with all of the above. Casual thoughts and observations are one thing, but if you're going to argue your point vehemently, show your work.

Moving on: I figured out the error in the KristenAndSara sequence; it's so noisy but quantized so flat that the psy-rd is thrown off and tries to pull that noise energy from wherever it can, in this case the hair. psy-rd 1.0 is way too much; it's actually trying to match MORE noise energy than exists in the original area, and totally mispredicting. It's much less ugly with 0.2 or 0.4, although it's apparent that psy-rd is also just a bad idea with very low bitrates. (Makes sense.) The effect completely disappears with --nr 100 even at --psy-rd 1.0.
Psy-RD is not done yet. It works at low strength (.3 or .4), but you will definitely see artifacts if you use higher strength, and we're not ok with that at all. If you're paying close attention to the devel mailing list you'll see that there are patches in the works. They won't be committed until we're sure they represent two steps forward and no steps back. Psy-RD is not the only visual quality improvement we're working on.

[LigH]

Plus, you are obviously not only working on (psycho-) visual enhancements, but also on speedups by making the code more efficient. Not many patches will have a noticable effect on their own, but their sum will probably be convenient.

:thanks: Go on, you are still not running out of good ideas.

[Atak_Snajpera]

I always give links to encoded files with command line. Screenshots are good for showing lack of details particularly mostly in dark scenes. With improved psy we can atleast quickly notice some progress. No idea why some of you switch to auto defence mode whenever I show weakness of current version of x265. At least end user knows that it is too early to switch from x264 to x265. I wouldn't waste my time for those tests if I didn't see posts claiming that x265 already offers better quality than x264. BTW. When I say better quality I mean more details at certain bitrate. I still have that marketing bull saying "Same quality (level of details) ,half bitrate" in my head ;)

[LigH]

Links to uncompressed sources are useful too for other people to recreate your results, and to compare against the original. Many uncompressed sources are available via the Xiph Media archive, but not all. The talking people in front of the dark blue background are hosted on a finnish server; x265 mentioned the exact filename of one of three in the VideoHelp forum so I was able to find it, but I am not certain if revealing the URL is allowed in public.

The file is called Johnny_1280x720_60.yuv It's 829,440,000 bytes. I can't provide a link... but you may be able to find a copy somewhere.

[destanig]

Dear Atak_Snajpera,

The feature that controls encoder's attention to fine detail is called Adaptive Quantization. Psy-rd is not about that and no amount of it will make x265 or x264 retain more details.

I encoded ~15 seconds of Drive with medium and placebo presets in the ABR mode. For the medium encode I used --preset medium --rd 6 --psy-rd 0.45 --aq-mode 1 --aq-strength 1.8. The bitrate is 1034 Kbps.
The clip is here:
https://mega.co.nz/#!4M8jWbSR!lSnm0u33yKTD7neJUXN73Txjt5kvNy3huriDEBGw1qM

For the placebo encode I used --preset placebo --psy-rd 0.4 --aq-mode 1 --aq-strength 1.8. The bitrate is 1045 Kbps. The clip is here:
https://mega.co.nz/#!QENxhCYa!LAndZQoBuXfe6pYxeQFOjFF26IF0l11Bjq0OipDnVmU

Both encodes look much better than your x265 and x264 encodes. They retain more details and lack the blocky noise of x264. What your tests show is that the default parameter tuning of x265 is worse than that of x264.

[mandarinka]

You can't say that in such a blanket matter. Both these features work together to preserve detail, textures and grain, as you can see in x264.

Using just AQ as a tool, you won't get grain retention, or textured flat areas. You need PsyRDO too, to compensate for the encoder's tendency to smooth and discard complexity.

[destanig]

The doc doesn't say much about PsyRDO. I could very well be wrong but my understanding is that it doesn't code any actual detail from the source at all. It looks at what details have been quantized out and adds noise in place of those to compensate for their absence.

On the other hand, AQ is exactly about removing bits from complex areas and using them in less complex areas like in flat textures.

[LoRd_MuldeR]

The doc doesn't say much about PsyRDO. I could very well be wrong but my understanding is that it doesn't code any actual detail from the source at all. It looks at what details have been quantized out and adds noise in place of those.

Video encoding is all about making choices! The standard RDO (rate distortion optimization) algorithm tries to minimize the "distortion" (in terms of pixel differences) as well as the "rate" (in terms of bit cost). So, for example, if a particular choice would result in very small bit cost, but also cause very high distortion, then the encoder won't make that choice. Similarly, if a particular choice would result in very small distortion, but cause extraordinary bit cost, the encoder won't make that choice either. In other words, RDO is all about finding the "best" compromise between the distortion and the bit cost. Unfortunately, a "small distortion" alone doesn't necessarily mean that the image looks "sharp" and "detailed" to the human eye!

Now what Psy-RDO does differently is that it considers one additional property: the "complexity" (in terms of variance). Consequently, with Psy-RDO enabled, the encoder will now prefer choices that keep the complexity (variance) of the coded picture as close as possible to the complexity of the original, even if that means a somewhat higher distortion and/or a somewhat higher bit cost. And the Psy-RDO strength controls how much weight the "complexity" measure gets in the encoder's choice making. It should also be obvious that setting the Psy-RDO strength too high will allow for very strong distortions and thus result in nasty artifacts. Keeping the "complexity" as close as possible to the original alone doesn't help either.

See also:
http://x264dev.multimedia.cx/archives/37

[LigH]

From a certain point of view, "noise" is also "detail"; it just depends if it is recognized as annoying, especially when chaotic and irregular. Or "overly regular" in the wrong situation.

[Atak_Snajpera]

Dear Atak_Snajpera,

The feature that controls encoder's attention to fine detail is called Adaptive Quantization. Psy-rd is not about that and no amount of it will make x265 or x264 retain more details.

I encoded ~15 seconds of Drive with medium and placebo presets in the ABR mode. For the medium encode I used --preset medium --rd 6 --psy-rd 0.45 --aq-mode 1 --aq-strength 1.8. The bitrate is 1034 Kbps.
The clip is here:
https://mega.co.nz/#!4M8jWbSR!lSnm0u33yKTD7neJUXN73Txjt5kvNy3huriDEBGw1qM

For the placebo encode I used --preset placebo --psy-rd 0.4 --aq-mode 1 --aq-strength 1.8. The bitrate is 1045 Kbps. The clip is here:
https://mega.co.nz/#!QENxhCYa!LAndZQoBuXfe6pYxeQFOjFF26IF0l11Bjq0OipDnVmU

Both encodes look much better than your x265 and x264 encodes. They retain more details and lack the blocky noise of x264. What your tests show is that the default parameter tuning of x265 is worse than that of x264.
btw. HERE IS MY 10 MIN DRIVE SAMPLE FOR DEVS.
https://mega.co.nz/#!5I1VzBxJ!12Q5FNeTscx9whhCmn922mlYnLYOYO_gktdJtT_7IXo

[Sulik]

In other words, RDO is all about finding the "best" compromise between the distortion and the bit cost.

IMO, the real problem is that we don't have a good measure of "Distortion", so RDO in its current form would be more appropriately called "RPO", or "Rate-PSNR-Optimization"...

[Daemon404]

In a literal sense he is correct. Psy-RD does *not* preserve detail. It is psychovisual, and in a simple sense, it recreates the complexity (http://forum.doom9.org/showthread.php?t=138293) of the source frame. That is not the same thing as preserving detail. Adaptive quantization, however, does, since well, it affects the quantizer! He may have worded it poorly, but he is not wrong.

IMO, the real problem is that we don't have a good measure of "Distortion", so RDO in its current form would be more appropriately called "RPO", or "Rate-PSNR-Optimization"...

Currently x265 is split. Some of the RDO has psy taken into account, some doesn't. The thing right now that is causing the most 'loss of complexity' is RDOQ, which currently is just using the SSE.

[LoRd_MuldeR]

IMO, the real problem is that we don't have a good measure of "Distortion", so RDO in its current form would be more appropriately called "RPO", or "Rate-PSNR-Optimization"...

And that's exactly why Psy-RDO has been created, I suppose.

One can also think of Psy-RDO as an enhanced distortion measure that takes into account "complexity" retention - in addition to the classical pixel-wise error (PSNR).

[Sulik]

And that's exactly why Psy-RDO has been created, I suppose.


It's definitely a step in the right direction, but I guess I'm not as confident in the existing various variance-based psy models. Seems like it's an area that has a lot of room for improvement but seems like little changed since the MPEG-2/TM5 days, in comparison to improvements in coding efficiency -> it seems like there should be both spatial and temporal considerations when measuring distortion (unless we consider residual rate itself a measure of temporal complexity ?).

[Sagittaire]

And that's exactly why Psy-RDO has been created, I suppose.

One can also think of Psy-RDO as an enhanced distortion measure that takes into account "complexity" retention - in addition to the classical pixel-wise error (PSNR).

yes Psy-RDO can be HVS compromise between different metric like PSNR, SSD or other.

[destanig]

Looks like my understanding of Psy-rd was oversimplified and may be it still is.

So if a certain detail is quantized out, PsyRDO tries to recover the lost energy by distorting the image.
The quantizer focuses on important details and the PsyRDO masks the loss of unimportant ones by introducing distortion that attempts to keep the original "complexity". This may work very well with things like grain at constrained bitrates: the quantizer doesn't code much of it and the PsyRDO just adds similar amounts of easy-to-code noise that looks like grain.

However, if the problem is with important details being quantized out, it's the quantizer that should be tweaked, not the Psy-rd. Increasing the Psy-rd setting is just going to add more noise in place of those details.

The problem with Atak_Snajpera's Drive encodes is lack of detail in flat areas, which is precisely the problem AQ was made to solve. Increasing AQ strength makes a huge difference here. And this is the point that I'm making with my encodes.

[LoRd_MuldeR]

So if a certain detail is quantized out, PsyRDO tries to recover the lost energy by distorting the image.

I wouldn't say it actively distorts the image.

Instead, with Psy-RDO enabled, the encoder will be biased towards making choices that preserve the energy/complexity, even if (but not necessarily) those choices result in a somewhat stronger distortion.

And here "distortion" refers to the classical distortion measure, like PSNR!


You can also think of it like this: If the original block had a lot of detail/noise, then Psy-RDO prefers a coded block that also has a lot of detail/noise, even if the detail/noise pattern isn't quite the same.

This is justified by the idea that the human eye has no idea how the original looked like. So detail retention can be more important than resembling the original as accurately as possible.

Of course this has its limits. If you set the Psy-RDO strength too high, the coded picture will diverge too much from the original and you are going to get other nasty problems than just "blurring" ;)

[mandarinka]

I wouldn't say it actively distorts the image.

Instead, with Psy-RDO enabled, the encoder will be biased towards making choices that preserve the energy/complexity, even if (but not necessarily) those choices result in a somewhat stronger distortion.

I think that is a bit unfortunate way to say about it, because it suggests that the "lower distortion" without psyrdo was inherently a better thing.

However problem is that the rate-distortion metric used is very limited in its "decision-making" ability. Even if plain RDO is supposedly shooting for "least distortion", its results are unbearably smoothed, because the metric doesn't care about noise-like features of the picture and instead supports smoothing them out.

When you look at many such encodes, it is painfully obvious that the smoothing/washing out is very strong and that it outweighs any gains of "decreased distortion" from absence of PsyRDO. In other words, PsyRDO increases distortion as meassured by SSIM and PSNR, but at the same time greatly decreases subjective (or how to call it) distortion caused by all the smoothing effects.

Hence I think it is not very useful to describe psyrdo as something that "increases distortion" or "makes up detail". I would say that ideally, it actually aims to do the same thing as plain RDO, only it has different criteria (not just psnr score)...

[Sagittaire]

The problem with Atak_Snajpera's Drive encodes is lack of detail in flat areas, which is precisely the problem AQ was made to solve. Increasing AQ strength makes a huge difference here. And this is the point that I'm making with my encodes.

AQ and psyrdo are not the same thing. If codec have problem with detail retention, you can't really solve that with AQ. AQ is here for better bits repartition (HVS sens) in picture with HVS masking like Luma, complexity or contrast. AQ can solve local problem in picture with better quality in dark/flat area for exemple. Anyway AQ can't solve problem to general grain retention in complete picture.

[x265_Project]

In a literal sense he is correct. Psy-RD does *not* preserve detail. It is psychovisual, and in a simple sense, it recreates the complexity (http://forum.doom9.org/showthread.php?t=138293) of the source frame. That is not the same thing as preserving detail.
Daemon404 - I'm pretty sure that you understand how x265 works, but I think your explanation may be confusing to some. I'll try to clear up the confusion in this thread.

Psy-RD is a simply a metric used for mode decision (the algorithm in x265 that decides how to encode a particular block). Psy-RD is off by default, but when it is turned on it is used to calculate the PsyRDcost in place of the normal rdcost for each candidate. PsyRDcost adds one metric to the mode decision - the psycost (the difference in energy between the source block and the reconstructed block). Psy-RD will bias mode decision towards candidates with energy (level of detail) in the reconstructed (decoded) block that more closely matches the energy of the source block.

Daemon404- When you said that Psy-RD "recreates the complexity of the source frame", some may have interpreted this as x265 modifying pixels (distorting the image). Except for a handful of obvious exceptions (noise reduction, dithering of high bit depth to lower bit depth, quantization, loop filtering) x265 doesn't modify the source pixels. Psy-RD is just a different method of choosing the best way to encode each block.
We're still working to adapt Psy-RD for HEVC, and to improve the Psy-RD algorithm. Our development team decided to make an early implementation available to get some feedback, but more changes are coming.

[Daemon404]

Daemon404- When you said that Psy-RD "recreates the complexity of the source frame", some may have interpreted this as x265 modifying pixels (distorting the image).

It was pretty clear from the post I linked that it doesn't modify the source pixels, IMO. Also considering the 'RD' in Psy-RD stands for rate distortion, it should also be somewhat clear, at least to anyone with a background in digital video.

[foxyshadis]

I always give links to encoded files with command line. Screenshots are good for showing lack of details particularly mostly in dark scenes. With improved psy we can atleast quickly notice some progress. No idea why some of you switch to auto defence mode whenever I show weakness of current version of x265. At least end user knows that it is too early to switch from x264 to x265. I wouldn't waste my time for those tests if I didn't see posts claiming that x265 already offers better quality than x264. BTW. When I say better quality I mean more details at certain bitrate. I still have that marketing bull saying "Same quality (level of details) ,half bitrate" in my head ;)

People get annoyed with you because you are demanding x265 put up or shut up over something they've never claimed. They don't even claim to be better than x264 at all, they just let the encoder stand for itself and announce its improvements, you're the only one who expects something more. It's the JCT-VC committee and the press that claimed "same quality at half bitrate", when everyone knows that was the design goals, not even a comparison of the JM with the HM, although that did get close. x264 is not JM, obviously.

You need to tone down the emotion or leave this subject alone. It's obvious that it's developed into a grudge for you, when the other posters here just want to experiment and encourage its development.

[xxxxx]

I don't understan how can people say x264 is better. I just made a test with a Big Bang Theory episode at 900kbps, x264 is terrible (people like vax, almost no details, unwatchable), while x265 gives almost the same quality like the source.
It will be perfect with 2 pass encoding, i really hope it's coming soon:)

[LigH]

A claim without proof and details, impossible to evaluate its meaning. And 1-pass bitrate based encodes provide worst quality distribution anyway.

[Kurtnoise]

It will be perfect with 2 pass encoding, i really hope it's coming soon:)
patches for that are coming (https://mailman.videolan.org/pipermail/x265-devel/2014-June/004554.html)...

[Atak_Snajpera]

I don't understan how can people say x264 is better. I just made a test with a Big Bang Theory episode at 900kbps, x264 is terrible (people like vax, almost no details, unwatchable), while x265 gives almost the same quality like the source.
It will be perfect with 2 pass encoding, i really hope it's coming soon:)

At least I post command lines and samples... Why should I treat your post seriously? Upload source sample like I did and then we will continue this discussion.

[kentaru]

People who are having difficulty with detail retention in their x265 encodes may want to try lowering the the max CU size. In my tests --ctu=16 works best for SD content and --ctu=32 works well for 1080p. The effect of lowering this value is far more pronounced than psy-rd.

[x265_Project]

Another psy-rd patch was committed, but more changes are pending. Don't get too excited about the first patch... better to wait for the additional patches that are coming. Hopefully another patch tomorrow, and further patches in the coming days.

[Sagittaire]

I don't understan how can people say x264 is better. I just made a test with a Big Bang Theory episode at 900kbps, x264 is terrible (people like vax, almost no details, unwatchable), while x265 gives almost the same quality like the source.
It will be perfect with 2 pass encoding, i really hope it's coming soon:)

well it's simple: simply because the test use particular source and/or particular bitrate (really low compexity, really dark, low quantizer encoding ... etc etc). With these source or/and profil encoding, psy optimisation in x264 work very well. Anyway in general case, and particulary at low bitrate, x265 produce really better result than x264, and by far. x265 is abble to encode complete movie at 720p in CDR size at acceptable quality level, not x264.

[Sagittaire]

A claim without proof and details, impossible to evaluate its meaning. And 1-pass bitrate based encodes provide worst quality distribution anyway.

crf mode produce exactly the same output quality than multipass mode ... without bitrate predictibility anyway.

[vivan]

well it's simple: simply because the test use particular source and/or particular bitrate (really low compexity, really dark, low quantizer encoding ... etc etc). With these source or/and profil encoding, psy optimisation in x264 work very well.Low complexity? Yeah, crowdrun and parkjoy have really low complexity.
And what is "profil encoding"?

crf mode produce exactly the same output quality than multipass mode ... without bitrate predictibility anyway.Since when 1-pass is multipass?