SubME 9 & Trellis 1 vs SubME 7 & Trellis 2

[zagnutty]

My first post here on this wonderful site.

I've recently been trying to save some of my HTPC storage space by re-encoding my 1080 movies to 720p. I've done the research, looked up what all the settings do. With all of my preferred encoding settings saved, I'm at my last barrier.

What kind of quality difference could one expect between using SubME 7 and Trellis 1 compared to SubME 9 and Trellis 2? Are we talking between 1-3%? Or something more like 10-15%? Believe me, I've searched high and low, but can't seem to find anything recent saying what kind of difference one can expect between Trellis 1 and 2.

SubME 9 and Trellis 2 are out of the question due to extensive encoding times. If I only had 1 or two movies, probably. But I have much more than that to encode. No animated material, strictly movies. Encoding using 2 pass average size of a DVD5.

[Dark Shikari]

The presets already provide you with good tradeoffs between low-level encoding settings; it's generally pointless to worry about the specifics beyond that.

[zagnutty]

The presets already provide you with good tradeoffs between low-level encoding settings; it's generally pointless to worry about the specifics beyond that.

I know. I've seen that said many times. But I'm one of those stubborn idiots that insists on tweaking everything. So can we just speak hypothetically then? Lets take subME out of the equation since my main question is really about what kind of improvement one can expect using trellis 2 over trellis 1. All I seem to find is vague answers saying "Use Trellis 2, it's better". But how much better? 3% better quality and compression, or 15%?

And I choose to use xvid4psp due to its flexibility and compatibility with various input formats. Unfortunately, it doesn't have the same type of presets one would find using one of the more preferred front end GUI's.

[Blue_MiSfit]

Feel free to test.

AFAIK, Trellis isn't considered a psychovisual optimization, so you can do (valid) SSIM tests etc...

For example:


x264.exe input.avs --crf 18 --preset slow --tune ssim --ssim --output output_trellis1.mkv
x264.exe input.avs --crf 18 --preset slow --trellis 2 --tune ssim --ssim --output trellis2.mkv


Adding the --ssim switch will report the metric at the end. You can see how much of a compression advantage it offers, and at what speed cost. I doubt you'll find it worth it. --preset slow includes --trellis 1 and --subme 8. The next "better" preset (--preset slower) has both trellis 2, and subme 9.

Have fun...

Derek

[Dark Shikari]

subme and trellis 2 aren't independent because trellis 2 operates on the modes selected by RDO, which are determined by subme.

[CruNcher]

thats where Adaptive Deadzones would come in they are RDO independent aren't they ?

[LoRd_MuldeR]

"Use Trellis 2, it's better". But how much better? 3% better quality and compression, or 15%?

What do your percent numbers refer to?

3% better quality? Then how do you express quality as a number? Or 3% lower bitrate for the same quality? Then how is "same quality" defined? :confused:

If, for example, you want to detect a bitrate saving as small as 3%, you would need a very precise "same quality" measure.

After all, you can do a 2-Pass encode with the same target bitrate and, except for Trellis, identical settings. Then judge which one looks better for your eyes!

And then you can decide whether the improvements in quality (if any) is worth the additional encoding time for you...

(Also keep in mind that the result will be bitrate-specific. If the test bitrate is chosen high enough, you certainly won't be able to spot a difference)

[looney]

What kind of shrinkage do you want and at what expense?

If you have really high quality 1080p (not sub 840p) encodes and you want to preserve it just for watching i'd recommend you splitting into few parallel threads and fft3d filtering (~2fps per 3.0ghz k10.5 core) into intermediary files and then use ffms2 to join all files before final encode. And reencode to some standardQ-1080p @5.5Mbps (10b) rather than extraHQ/superHQ720p at same 5.5M bitrate.

On behalf subme 7 vs subme9 i'd say that if you want some dumb LQ encode you should use subme8 in 2nd pass. I personally use ultraslow subme10 if the source is really pristine and rich in details, while subme8/9 is good for some cheap encodes as it is 25%/20% faster vs subme10.

If you really want to have fast encodes then you skip fft3d filtering and use hex instead umh it's some where subme8 is some 30% faster than subme10 (x4@2.5G ~5.8fps).
Disadvantage of hex is somewhat bigger files and reduction in picture smoothness/cohesion, it fills with more ringing granularity effect on 720p encodes <3.2Mbps

If you use trellis you'd get nothing on hex but it umh really complements umh encodings (i use trellis 2 in all 2nd passes)

I hope i somewhat help you.

So can we just speak hypothetically then? Lets take subME out of the equation since my main question is really about what kind of improvement one can expect using trellis 2 over trellis 1. All I seem to find is vague answers saying "Use Trellis 2, it's better". But how much better? 3% better quality and compression, or 15%?

I don't think sou should expect 15% better compression with trellis. But you seem to deliberately forget the mst important part of mpeg4 ME hex vs. umh makes much sense with trellis 2 enabled, other way you should use hex w/o trellis and use high bitrates and subme7 for faster but larger encodes. If you're on quality and small size than your encoding sould be pre-filtered and x264 encoding itself will be at least 4times slower if you target low bitrates at almost same quality but with umh+sumbe10+trellis2. Trellis itself wont make picture magically better

Otherwise forget about trellis imho if you can afford 12-18Mbps. I'd rather use that extra bitrate for 1080pSBS :wink:

Trellis changes picture, but most of us humans cant recognize that when watching on some 42" from 2.5m distance so it's better to use it but only to be complement for insanely slow motion estimation and lower bitrates.

[Bloax]

fft3d filtering (~2fps per 3.0ghz k10.5 core)
Why wouldn't you use FFT3DGPU, since that ought to be both faster and give x264 more CPU power (thus overall faster).

[Blue_MiSfit]

You probably should, unless you're encoding on a server without a fancy-pants GPU, or need to access your machine via remote desktop.

Derek