SVT-AV1 v3.1.0

[charliebaby]

SVT AV1 v3.1.0.113 Build Windows .EXE
SVT [version]: SVT-AV1 Encoder Lib v3.1.0.113
Svt[info]: SVT [build] :CCCG 15 Win-64 bit

https://www.mediafire.com/file/glx3of0xvxndka9/SvtAv1EncApp+3.1.0.113.7z/file

[CruNcher]

wow not so bad transcoding VC1 HVS results with this new SVT-AV1 3.1.0 @ P9 CRF 21 VQ
P10 CRF 21 VQ/PSNR/SSIM has a bug it seems with fade complexity can cause strange huge macroblocks in some WMV that coming out of fade sequences.
But P9 by default seems very stable and still fast.

Lets test Tall Ship (searching databse)

Really impressed by the Performance of skin retention as well freckles, pores, small moles and veins shine through at high speed.
The ranges where you need medium x264 to retain high frequency details and that much much slower.

Nice Optimization it feels like a VC1 successor but easily beating it and X264

Good old VP3.2 times :)

and the overall fixing of some WMV still retains the high frequency nicely even after the inloop force fixing in faster scenes (temporal high frequency retention).

MP@HL WMV3
Advanced@L3 WVC1

nice results by default for the transcoding performance, it either makes it smaller or hits the size.

The Fusion of Daala works out in the end.

[GeoffreyA]

By the way, the new version of the fork was also released last month and has some useful additions over mainline SVT-AV1.

https://github.com/BlueSwordM/svt-av1-psyex

[CruNcher]

i tried nekotrix his essential branch over his handbrake implementation but it totally kills the Performance advantage hes tuning with (performance ratio doesn't seem to matter her clean waifus do)

P10 causing immediately this fade bug (first experience of it) shows that at CRF 21 with Adaptive Quantization (VQ Implementation) it can't sustain it i see this problem for now as a failure and i will try to tune it out.

P9 so far is very unique in terms of Performance/Compression/Result/Stability Ratio with just VAQ and stands pretty much on its own

10,9,8,7,6 are the very interesting regions of performance ratio vs X264 and Atemes Encoder.

thank you im aware of the Psy branch still going to see where the tuning goes towards blueswords and the psychovisual expert group, remembers me of how work was done on it together for X264 bringing in VAQ (Low Performance Impact) and PSY-RD (Higher Performance Impact + ME Precision).

But also reducing the Performance and this needs to be carefully evaluated against P9 awesome PSY/Performance/Compression/Stability ratio that stand on it's very own for now very advanced MPEG-4 ASP PSY ratio.

i was a bit shocked how SVT-AV1-Essential (nekotrix) was cutting away from that i think it was by -~3x slower in the handbrake mod if i remember that fast test i made right with that mode that nekotrix named (Faster) not sure how that fits vs P9 default tough calling it that but maybe introducing 2 separate ranges in terms of PSY ME combination would be a interesting idea.

The slower adaption into Mainline looks like the Main devs are very carefully considering the Performance impact throughout the HVS space and possible patent issues of course.


SideNote: Interesting how https://svt-av1-psy.com/video/ fails function properly in its interop on some older chromium Blink browser but works flawless on Firefox Gecko.


Im combining AV1 a lot with AI outputs and im impressed of the combination
on a 8 Core Ryzen 3700x x264 has hardly a performance chance

i still didn't experiment with sharpness yet pure default P9 @ CRF 21 very impressed of the optimization.

Also in non live streaming input transcoding use cases i have to ask a very severe question and evaluate it in terms of 2-pass efficiency vs a -3x slower 1-pass psy branch tuning.

i did some nice 50% testing of VC-1 to AV1 at super high speed and it looks really good i love testing first with P9.

[GeoffreyA]

I haven't tried Nekotrix's builds, but regarding Blue's PSY fork, I haven't had good results yet. There are many settings to tune and it is a bit overwhelming, knowing which knob does what.

For my part, I'm fed up with all these new codecs, AV1, VVC, the works. I wish H.264 had better compression.

[CruNcher]

you wont get much more out of it but performance on the encoder side could be surely still improved here SVT-AV1 shows real teeth now @ 10,9,8,7
but some of that speed comes from the absence also of b-frame complexity on the other side here P10 can fail really heavy on its nose it seems but i still have to analyze this fade in disaster as AV1 should even handle this better im surprised it failed on this test case but only at P10 producing artifacts on the VAQ layer coming out of a fade that has immediate high motion and high frequency response.

[Leo 69]

you wont get much more out of it but performance on the encoder side could be surely still improved here SVT-AV1 shows real teeth now @ 10,9,8,7
but some of that speed comes from the absence also of b-frame complexity on the other side here P10 can fail really heavy on its nose it seems but i still have to analyze this fade in disaster as AV1 should even handle this better im surprised it failed on this test case but only at P10 producing artifacts on the VAQ layer coming out of a fade that has immediate high motion and high frequency response.

Which AV1 encoder would you recommend for maximum visual fidelity? So many forks popped up lately, it's a bit overwhelming. You seem to have tested them a lot.

[CruNcher]

Not yet im in the middle of finding out what causes this fade in artifact issue and testing between the release and Psy version on P10 currently so far i would prefer P9 21 CRF @ 3.1.0 as primary testing because of its crazy speed transcoding and stability.

but i see definetly improvements in the old psy here but higher bitrate of course but p10 result is stable this way so bitrate maybe starves for the fade in to be 100% artifact free i guess a bit in 3.1.0 release P10.

this works with the fade in being stable and artifact free

Svt[info]: -------------------------------------------
Svt[info]: SVT [version]: SVT-AV1-PSY Encoder Lib v3.0.2-A
Svt[info]: SVT [build] : Clang 21.1.0 (https://github.com/llvm/llvm-project.git 3623fe661ae35c6c80ac221f14d85be76aa870f1) 64 bit
Svt[info]: -------------------------------------------
Svt[info]: Level of Parallelism: 5
Svt[info]: Number of PPCS 76
Svt[info]: [asm level on system : up to avx2]
Svt[info]: [asm level selected : up to avx2]
Svt[info]: -------------------------------------------
Svt[info]: SVT [config]: main profile tier (auto) level (auto)
Svt[info]: SVT [config]: width / height / fps numerator / fps denominator : 1280 / 720 / 30000 / 1001
Svt[info]: SVT [config]: bit-depth / color format : 8 / YUV420
Svt[info]: SVT [config]: preset / tune / pred struct : 10 / VQ / random access
Svt[info]: SVT [config]: gop size / mini-gop size / key-frame type : 305 / 16 / key frame
Svt[info]: SVT [config]: BRC mode / rate factor : CRF / 21.00
Svt[info]: SVT [config]: AQ mode / Variance Boost strength / octile / curve : 2 / 2 / 6 / 0
Svt[info]: SVT [config]: sharpness / luminance-based QP bias : 1 / 0
Svt[info]: SVT [config]: temporal filtering strength : 1
Svt[info]: SVT [config]: QP scale compress strength : 1.00
Svt[info]: SVT [config]: Noise Normalization Strength : 1
Svt[info]: SVT [config]: Keyframe TF Strength : 1
Svt[info]: SVT [config]: PSY-RD Strength : 1.00
Svt[info]: SVT [config]: spy-rd : non



much i cant say so far im in the middle of testing this right now

this fails on 3.1.0 release (default setup for P10) artifacting on fade in

Svt[info]: SVT [config]: main profile tier (auto) level (auto)
Svt[info]: SVT [config]: width / height / fps numerator / fps denominator : 1280 / 720 / 30000 / 1001
Svt[info]: SVT [config]: bit-depth / color format : 8 / YUV420
Svt[info]: SVT [config]: preset / tune / pred struct : 10 / VQ / random access
Svt[info]: SVT [config]: gop size / mini-gop size / key-frame type : 305 / 16 / key frame
Svt[info]: SVT [config]: BRC mode / rate factor : CRF / 21
Svt[info]: SVT [config]: AQ mode / Variance Boost : 2 / 0
Svt[info]: SVT [config]: sharpness / luminance-based QP bias : 0 / 0


with the old psy version i can definitely say i see a even larger skin retention high frequency improvement but bitrate is higher so its hard to say if it is better or not i would first like to fix the fade in before concluding anything else.

a visible interfering stability issue has priority to find a fix spot before even going at the retention quality.

i can see these huge macroblocks appearing for maybe just 2 frames inside the fade in but its psychovisual disturbing anyways and has absolute fix priority its a no go to happen.

i first gonna make a test encode witch replicates the psy result size with 3.1.0 to see the CRF deviation and will watch if the fade in problem reoccurs at the same size on the psy branch setup or maybe i gonna straight target the P9 SVT-AV1 Release size result at SVT-AV1-PSY P10.

i would predict the CRF difference to be CRF 24 vs CRF 21

the overall transcode time of those 23 minutes is just 2:14 mins so that is super fast to test through.

yes and it seems my prediction target was even correct +3

ah no almost exactly 24.5 so +3.5

ok on SVT-AV1-PSY Encoder Lib v3.0.2-A it artifacts now on i think 1 frame before the fade out even starts crazy on P10 at almost the same size as P9 CRF 21

to complex to word this all out better in table form

so at the almost same target size neither the Psy Encoder stays 100% stable on the fade sequence @ P10 but overall better in terms of the placing before the fade out and not inside the fade in though we can argue now if that is psychovisual better i guess no normal human would have seen the artifacting in the fade in as i did but shortly before the fade out i guess lot more would realize it.

So just psychovisualy i would assume the. error to better stand inside the fade in

the easiest way now would be saying P10 is unusable for this complexity goal so just move towards P9 and be happy before other problems hunt you.

though i would like to know now what the preset differences are between P9 and P10 internally and how exactly they affect performance

So yeah the artifacts moved from the fade in into short before the fade out but you could say they happen in the fade out now.

so the scene buildup is half body closeup talky talky into camera->fade out->fade into action (temporal movement)

and this is quiet problematic for P10 to get stable without artifacting either in the fade out or the fade in sequence.

but in the psy branch @ P10 CRF 24.5 its just 1 frame that fails in the fade out just 1 freaking frame ah no its more its hard to spot this exactly you have to seek frame exact to the exact point its several frames artifacting though in the fade out its on a white wall moving into black in the fade in it would be much more problematic in terms of ROI hit.

the white wall has a lot of high frequency details and some regions blur out as well
so its pretty much exactly 3 frames when you don't count the blurry blocks as part of the real error

Frame 1/2/3/(4 Fadeout)

SVT-AV1-PSY Encoder Lib v3.0.2-A P10 CRF 24.5 Fade out error progression

https://i.postimg.cc/h4QB7Ttd/fade1.pnghttps://i.postimg.cc/wTF0qbYc/fade2.pnghttps://i.postimg.cc/bvRvWbtJ/fade3.png
https://i.postimg.cc/2SSbY0CW/fade4.png

i cant show the SVT-AV1 3.1.0 P10 CRF 21 Fade in error progression for several reasons but it's right after the fade out here finishes and the fade in sets in ;)


i tested this now the error progression is a bit different and inside the fade out and fade in but bitrate is further reduced

Svt[info]: -------------------------------------------
Svt[info]: SVT [version]: SVT-AV1 Encoder Lib v3.1.2
Svt[info]: SVT [build] : GCC 13.2.0 64 bit
Svt[info]: -------------------------------------------
Svt[info]: Level of Parallelism: 5
Svt[info]: Number of PPCS 76
Svt[info]: [asm level on system : up to avx2]
Svt[info]: [asm level selected : up to avx2]
Svt[info]: -------------------------------------------
Svt[info]: SVT [config]: main profile tier (auto) level (auto)
Svt[info]: SVT [config]: width / height / fps numerator / fps denominator : 1280 / 720 / 30000 / 1001
Svt[info]: SVT [config]: bit-depth / color format : 8 / YUV420
Svt[info]: SVT [config]: preset / tune / pred struct : 10 / VQ / random access
Svt[info]: SVT [config]: gop size / mini-gop size / key-frame type : 305 / 16 / key frame
Svt[info]: SVT [config]: BRC mode / rate factor : CRF / 21
Svt[info]: SVT [config]: AQ mode / Variance Boost : 2 / 0
Svt[info]: SVT [config]: sharpness / luminance-based QP bias : 0 / 0


but this means CRF can be even lowered now and lets see what then happens in both fades :)

hmm even with +50 mb total more then all the others it still bugs out in those fade out/in i guess now even reaching the source bitrate 1/1 it would still fail

~ 100 mb away from source size

so where we are now

591 MB (619.898.102 Bytes) (Source)
483 MB (506.938.708 Bytes) (Fade Stability Problems P10 CRF 19)
510 MB (535.372.923 Bytes) (Fade Stability Problems P10 CRF 18)
560 MB (587.225.340 Bytes) (Fade Stability Problems P10 CRF 17)
591 MB (620.247.715 Bytes) (Fade Stability Problems P10 CRF 16) Source Reached


No go with P10 so far getting this stable the source reach should be error free

ok so P10 is useless for this Source point blank it cant handle the fade complexity in 1 pass

now we compare very carefully the difference on P9 (no weighted prediction problems)

2:07 min 591 MB (620.247.715 Bytes) (Fade Stability Problems P10 CRF 16) Source Reached
2:16 min 619 MB (649.146.872 Bytes) (No Fade Stability Problems P9 CRF 16) Source Reached overflow
2:07 min 227 MB (239.019.651 Bytes) (No Fade Stability Problems P9 CRF 30) Compression Gain
2:12 min 434 MB (455.611.727 Bytes) (No Fade Stability Problems P9 CRF 21) Compression Gain
2:13 min 455 MB (477.947.766 Bytes) (No Fade Stability Problems P9 CRF 20) Compression Gain
2:14 min 501 MB (526.042.387 Bytes) (No Fade Stability Problems P9 CRF 19) Compression Gain
2:15 min 532 MB (558.178.566 Bytes) (No Fade Stability Problems P9 CRF 18) Compression Gain
2:16 min 584 MB (613.050.508 Bytes) (No Fade Stability Problems P9 CRF 17) Compression Gain


i didn't looked at metrics yet but i would say P9 CRF 21 is a good starting point to build many things up on
but surely we also want to know what it does @ 50% hitting a clean 300 mb here would be really nice lets see where handbrakes default 30 brings us to

My first Advise don't use P10 for anything that is more complex as to Render out the Fire Burning in the Carmine the whole night long.


Im doing my own Metric tests and evaluation rounds and also try to Performance Sync High Speed X264 with SVT-AV1

but metric wise no chance for X264 Performance vs P9 Performance it absolutely roast it on default setups.

actually i try to find the weakness doing visual/ and metric performance evaluation vs P9

but on the performance tuning side SVT-AV1 devs did magic with P9

P10 is mostly unusable in 1 pass at least (just to unstable for generic usage) it is solely optimized for Video Conferencing/Surveillance Cameras with static low complex camera setups only in normal directional movement and zoom (Remote Webcams).

P10 = Webcam Mode/Smartphone Live Recording (Video Conference Mode) it is not designed to handle artificial fade complexity nor any kind of eye blinking simulation (POV complex simulation or First Person simulation Game Scenario with complex shader effect designs).
P9 = Start of Generic System Latency Usability Mode

and i pretty much put my PSY experience contribute tuning on XviD/X264/ATEME PSY vs VMAF

https://i.postimg.cc/3rSbcmVr/metric2.jpg
https://i.postimg.cc/BQ2mPg16/metrics.jpg


P8/P7/P6

Nekotrix Essential builds are a very special use case scenario which fail on the generic use case side as he states himself

His work entirely concentrates on the upper level 10 bit HDR part and a specific Performance region and his optimization focus seems to be SSIMULACRA2.

on this side especially the PSNR Metric will shift quiet a bit

geeez i better go with my own piping pipeline handbrake is overall nice but with all these setups and build differences you use oversight of what really goes on in the background and the different builds can react weird on your custom preset up to the point another build wouldn't show it anymore

and interop on Windows NT 6.1 for no reason whatsoever fails

[CruNcher]

Some nice test content for testing AI Reconstruction Model/Encoder ME and Adaptive RC Transcode efficiency

behind the scenes combines multiple camera recording styles artificial post artifacts

https://www.youtube.com/watch?v=8i9q7lAfd-s

Main Video uses excessive film grain

https://www.youtube.com/watch?v=V9vuCByb6js

really awesome visual effects into old camera and media recording styles chromatic aberration effects, artificial resolution degradation SD 90s downscaling

https://www.youtube.com/watch?v=uCNr2iCgSPs

Some Volumetric Light,Global Illumination, Refelection, Focus (Blur Layer) Changes, Complexity non stroboscopic (static) Stage Setup

https://www.youtube.com/watch?v=xeLZz-ST0d8

Stroboscopic Eratic Level 1 (Low Epileptic Seizure Danger), Focus (Blur Layer) Changes, (Heavy Foundation Makeup)

https://www.youtube.com/watch?v=Em3cRUeOqcY

Wide Directional Strobes

https://www.youtube.com/watch?v=oJhsh4lxfGM


Awesome Visual Artistic Post Light and Cinematographic Team Behind this entire Production


PS: Geeez i find my encodings to huge for the quality but when I'm looking at some of these P2P Piracy Stream and Semi Pro FAST Stream encodings out there i would say nope everything perfect for 720p 80 min with AV1, how those pirates get it that the server choosing the wrong stream content i guess will stay a mystery hrhr ;)
on their deceived file sharing networks (mostly Russian War Finance Networks) they have systems to buy and automatically register domains and do it hourly its crazy and its fully automated and they buildup their own commercial shadow network infrastructure (Russian/Chinese Tor) cloaked as SEO optimization services (legit business entities) they fool so many into believing they are Pirates or greedy SEO Idiots but in reality they are Russian Assets (FSB/SVR) working under Control of Russian Agencies and they collaborating with Chinese (MSS) entities.

I wonder how many Smartphones in Europe are under Russian Agency Control embedded into this or better these Networks as they can adapt and split if necessary by now.

Be careful "Who your Cyberfriends really are"

Im not sure what hurts more anymore all this Child Abuse going on or all this Lying.


1080x still a 500 s buffer on 10 Mbits PowerSave Mode (Firefox) -20W


Crazy too look back and see how bad optimized Chromiums AV1 decoding was from the sart vs Firefox even with all that multithreading and latency advantages on the top render core seeing AV1 Decoding needing a own Power plant vs VP9 is crazy.