UK based DeepRender.AI Team

[CruNcher]

The First (announcing it) that want to disrupt the VPU/GPU with the NPU with a full working EDGE implementation challenging SVT-AV1 P7 with a highly optimized 3.x miillion (YoloV2 range) param model 150 mb int8 quantized

https://deeprender.ai/

i think all the tests where conducted with 0 PSY optimizations (by the Psychovisual Expert Group) neither vs SVT-AV1 + VB at that time

Though it is not correct to assume they are the first if there should be somewhere a hidden development project running which results get not published, neither vs known academic research.

Especially the possibility that Topaz Labs works for a long time now on it is very plausible, rather secured and who knows if Bytedance or Tencent doesn't also surprising us here.

So yeah we should prepare for a lot more announcements in that exciting EDGE disruption of Neuromorphic Computing on the edge to practical Quantum Computing :)

Their Presentation Slides have similiar CID to Elemental/Nvidia :D

they really use that green wtf give us familiarity with a top company that is so weak manipulation and realy 0 creativity in developing their own CID makes 0 sense as well when you call Nvidias GPU to weak and Copy their Presentation Style it made sense for Elemental running on CUDA exclusively it makes 0 sense for them.

Lot of compares are already outdated and obsolete since SVT-AV1 3.1.0

[GeoffreyA]

I haven't watched the comparison yet, but their website should tone down the marketing, especially since they're still behind VVenC, according to Jan. Does it solve the problem of grain?

[CruNcher]

honestly i predict they will at least reach it in short time
it is also highly in chinas interest to help them out here attacking others this disruption has a lot of benefits for them they will actively "hidden" push it as well
pretty complex what is going on currently, what really makes me feel sick how many human life's cease to exist for it in a very unnatural way.
And how inside this Chaos many are transforming or even get born into something you hardly would call human anymore.

[nhw_pulsar]

I haven't watched the comparison yet, but their website should tone down the marketing, especially since they're still behind VVenC, according to Jan. Does it solve the problem of grain?

Jan Ozer information are certainly solid, so if he says Deeprender codec, which is claimed to be the best practical end-to-end learnt video codec, is behind VVenC, then the next standard H.267 from MPEG will be certainly ECM, as submissions are closed now... Yes ECM incorporates now quite some NN processings, but I don't want to speak too quickly maybe H.267 will be an undisclosed super project that beats ECM -even if some tech companies' directors of research say that end-to-end learnt video codec will be ready rather for the next standards in 5 years-... we'll see :)

Cheers,
Raphael

[CruNcher]

Problem when china attacks they do not attack with 1 or 2 companies but 50 and as much student research power when you have 30 at deeprender you have 100*2 in the labs of Bytedance and Tencent collaborating over Tshingua and Singapore with others like Alibaba and their cloud and cloud mirror platforms exploding.
And the Datasets they gather are enormous also from combined data gathering and Research collaborations with Russia.
And you can never be sure how many Spy students are working for you, that might even be so far to execute your knowledge if it has to be.

[Z2697]

Is my network trash or their website just inaccessible/slow AF at the moment?

[Z2697]

Why somehow suddenly some "attack" from China and/or Russia is involved???
And not even in a metaphorical way, literally
many human life's cease to exist for it in a very unnatural way

Like what the hell, are we talking about video compression technology?

[nhw_pulsar]

Is my network trash or their website just inaccessible/slow AF at the moment?

No problem here, the website is completely OK. When I see their website, I find it nice that they could build up a nice team to develop their AI technology (compared to the hardship you have with wavelet technology...).

Cheers,
Raphael

[Z2697]

It's hard to believe that M4 with its 10 CPU cores can only encode SVT-AV1 preset 7 at 9 FPS but can encode VVenC preset faster at 10.7 FPS.
If you ask me, forget about quality, SVT-AV1 is crazy fast, good job well done, really. At least on x86-64 chips. I can encode SVT-AV1 preset 7 1080p YUV420P8 video at 35+ FPS on Ryzen 9950X single core (lp=1)!

22 FPS from a 38 TOPS NPU, if it's true, it's a killer feature. Decoding can be a bit of challenge tho.

[CruNcher]

You have to keep in mind

SVT-AV1 (v2.3.0)

forget about the non tech part i wrote it would get way to complex to understand how all this is connected.
And how these chain reactions of consequences work out in space time.

[Z2697]

Well if I don't have a SVT-AV1 v2.3.0 encoder binary lying around I might have believed that.
That being said, I really don't have a M4 Mac lying around, so I can't say anything about it.
Maybe the ARM optimization was basically none at 2024-10-28, I should go search around.

[GeoffreyA]

honestly i predict they will at least reach it in short time
it is also highly in chinas interest to help them out here attacking others this disruption has a lot of benefits for them they will actively "hidden" push it as well
pretty complex what is going on currently, what really makes me feel sick how many human life's cease to exist for it in a very unnatural way.
And how inside this Chaos many are transforming or even get born into something you hardly would call human anymore.

Film grain is the biggest problem in encoding, so if anybody can solve that, if indeed possible, grain being random and raising entropy, it will be an advance.

Regarding the other points, I don't understand. Are you referring to AI?

Jan Ozer information are certainly solid, so if he says Deeprender codec, which is claimed to be the best practical end-to-end learnt video codec, is behind VVenC, then the next standard H.267 from MPEG will be certainly ECM, as submissions are closed now... Yes ECM incorporates now quite some NN processings, but I don't want to speak too quickly maybe H.267 will be an undisclosed super project that beats ECM -even if some tech companies' directors of research say that end-to-end learnt video codec will be ready rather for the next standards in 5 years-... we'll see :)

Cheers,
Raphael

Yes, doubtless the next crop of codecs will have a lot of NN parts and there will be gains. Eventually, one could imagine whole scenes being generated from source metadata; but I think philosophy, rather than NNs, will advance video compression. When something can't be squeezed any more, what comes next? Current encoders have hit critical mass of concepts and the improvements are diminishing.

It's hard to believe that M4 with its 10 CPU cores can only encode SVT-AV1 preset 7 at 9 FPS but can encode VVenC preset faster at 10.7 FPS.
If you ask me, forget about quality, SVT-AV1 is crazy fast, good job well done, really. At least on x86-64 chips. I can encode SVT-AV1 preset 7 1080p YUV420P8 video at 35+ FPS on Ryzen 9950X single core (lp=1)!

SVT-AV1 is extremely fast and can beat x264 on some presets.

[nhw_pulsar]

Eventually, one could imagine whole scenes being generated from source metadata;


Yes, you're right, ultimately we could imagine that given a reference keyframe, the AI completely animates the keyframe with description texts/prompts, which could potentially drastically diminish bitrate. But certainly there will be approximations (for example in the motion) and quite some differences compared to the original source sequence. But do these differences will be acceptable and the AI-animated sequences pleasant to watch in the end? Maybe at first for some specific scenarios?

Cheers,
Raphael

[GeoffreyA]

Yes, you're right, ultimately we could imagine that given a reference keyframe, the AI completely animates the keyframe with description texts/prompts, which could potentially drastically diminish bitrate. But certainly there will be approximations (for example in the motion) and quite some differences compared to the original source sequence. But do these differences will be acceptable and the AI-animated sequences pleasant to watch in the end? Maybe at first for some specific scenarios?

That's the challenge. Perhaps each frame could be tethered to the original in a low-cost way, constraining the output, reducing differences. It might be acceptable for social media or YouTube content, but content preservation, perhaps not.

[benwaggoner]

I haven't watched the comparison yet, but their website should tone down the marketing, especially since they're still behind VVenC, according to Jan. Does it solve the problem of grain?
Everyone makes big promises until they hit anime and grain.

[CruNcher]

yeah i would like to see how a Beamr CABR SVT-AV1 fares against deeprender in a full practical service setup its the pretty maximum DCT VOD optimization

as Neuromorphic Processors going to flood also from China the markets and scaling much faster it seems a already a lost cause in terms of future improvements what you can do entirely on it on the EDGE.

We are in the transformation phase still but it's already the end of old paradigms and we surely gonna see much more also in the range of specialized high efficient purpose model.

Even Beamrs full GPU idea is somehow total nonsense when you can do everything on the NPU low powered with a smaller GPU core for just some generic operations to use so it makes much more sense to scale the NPU up and make the VPU space free terminating it as well.

I think even Nvidia will sooner or later going to obsolete the VPU entirely.

[Z2697]

Can you please... use some more punctuations?

[GeoffreyA]

Everyone makes big promises until they hit anime and grain.

Exactly. Anime and grain are the benchmark.

[nhw_pulsar]

Everyone makes big promises until they hit anime and grain.

Maybe grain will not be a main problem for AI? As, if the keyframe has grain, the AI, that animates the keyframe, should keep its grain then?

[CruNcher]

keeping the grain is not the major problem a smart enough AI model could replicate (mirror,parrot) the grain in a much more advanced way on the fly

only how you manage the resources and how you distribute it inside the system (hive) matters and how you balance the system in many ways not just technical (freedom/dignity/inclusion/behaviour/law/response/consequences).

Already a lot of resources get wasted because we still have not found a way to manage the digital economy efficiently,integrate it in a responsible way into the system and society so that no one is left behind chances equalized and power abuse also by the implementers controlled.

[ksec]

Worth reading this blog post.

https://deeprender.ai/blog/future-ai-codecs-specialisation

[benwaggoner]

keeping the grain is not the major problem a smart enough AI model could replicate (mirror,parrot) the grain in a much more advanced way on the fly
Yeah, it is a challenge to make grain that both matches the original aesthetic and is also efficient to encode. I've got a patent on the approach, but I'm not aware of anyone having done a practical implementation of it yet.

[benwaggoner]

Worth reading this blog post.

https://deeprender.ai/blog/future-ai-codecs-specialisation
Wow. The energy/compute costs of training a unique model per title are going to be substantial.

This is all interesting in theory, but I'm going to want to see some actual full titles done this way before offering an opinion on how it is in practice.

[nhw_pulsar]

Wow. The energy/compute costs of training a unique model per title are going to be substantial.

This is all interesting in theory, but I'm going to want to see some actual full titles done this way before offering an opinion on how it is in practice.

DeepRender seems to have really great technology, or I misread? They say that they can create a 300KB to 2MB specialized model for a 2-hours-long movie which will lead to a 50% bitrate reduction compared to VVC/ECM? I am certainly a big amateur in neural networks, but such a tiny size model doesn't sound like an exceptional feat?

[Z2697]

Perhaps they only need to customize a portion of the model, or the entire model might be small... given that it runs pretty fast on a "weak" NPU.

[CruNcher]

their model seems to be in the range of just for example the VAE Decoder/Encoder makes perfect sense there is no text encoder or much other complex logic we talk about 3.2 million params though they find a 150 mb file for encoding/decoding to much which is around 2x the size of a complex 3D Engine these days
in the range also of a complete FFMPEG static binary

[benwaggoner]

DeepRender seems to have really great technology, or I misread? They say that they can create a 300KB to 2MB specialized model for a 2-hours-long movie which will lead to a 50% bitrate reduction compared to VVC/ECM? I am certainly a big amateur in neural networks, but such a tiny size model doesn't sound like an exceptional feat?
In my long years in the industry, I've found that the more extraordinary the claims are for any new encoding technology, the less likely it is something of general utility.

The numbers sound cool, but I am going to reserve judgment until I can actually watch the results of some test content I know.

[nhw_pulsar]

In my long years in the industry, I've found that the more extraordinary the claims are for any new encoding technology, the less likely it is something of general utility.

The numbers sound cool, but I am going to reserve judgment until I can actually watch the results of some test content I know.

Hi Ben,

Yes I totally agree with your pragmatic approach, in the end anyone claiming great/exceptional compression technology should at least give their codec binaries.

For learned image compression, I am sorry if I tell stupidity as I am a total newbie, but we often read that we must have a training strategy, but this phrasing does not reassure me because I fear that you'll create a custom model that only works on the test/training images, and so total compressed size should be compressed files size + model size...

[CruNcher]

In my long years in the industry, I've found that the more extraordinary the claims are for any new encoding technology, the less likely it is something of general utility.

The numbers sound cool, but I am going to reserve judgment until I can actually watch the results of some test content I know.

yeah i would say we already reached a pretty saturated point under current possible physic space and things work out already in pretty efficient ways we are certainly not beg 2000 anymore.
it wont be easy to push this all over the edge with so many possible consequences.

When im streaming AV1 in pretty 4K on 10 Mbps without problems its already a WOW effect

when their is only a minimal perception change that wow effect would be still there but it really highly depends how much WOW it would be at that point anymore ;)

of course reduction of streaming services their costs still has some value but

just look where we really arrived

https://forum.doom9.org/showthread.php?t=186526

surely the reduction of costs is still value and surely for Low Latency Live Streaming we can still shift the improvement (especially valuable for the (adult) live streaming industry)


AI for me is primary interesting in a combined space workflow not just because of the general improvement but the data handling itself inside a very optimized system core namely where the NPU getting most of the resources inside of by deprecating the space of other functions.

[GeoffreyA]

What would be useful, too, is an AI model that worked on an encoder codebase, rather than video, taking an early, poor-quality version to a mature one. For example, putting in an early x264 and getting one from the 2010s. This would save a lot of time in encoder development.

[CruNcher]

When a AI goes to rewrite x264 in rust i gonna craze out

[benwaggoner]

When a AI goes to rewrite x264 in rust i gonna craze out
Wow. If we had an AI that was super good at scalar to SIMD conversion for multiple architectures, that was be amazing.

And startling!