Thread 106560011

Intel just revolutionized computer graphics with gaussian splats

Within like 3-4 years virtually every single real-time simulation or video game is going to be like 2x more efficient due to this tech

>>106560011 (OP)
>Within like 3-4 years virtually every single real-time simulation or video game is going to be like 2x more efficient due to this tech
You're way too optimistic about timelines desu

>>106560029
there's already a unity engine demo with gaussian splats that works flawlessly

>>106560011 (OP)
General quantum cumputing on every desktop and AGI in two more weeks.

more photo-realistic garbage, just what videogames needed

>>106560076
Are they out of the uncanny valley yet?

>>106560054
doesn't matter, most studios are insanely slow moving. I hope i'm wrong and you're right, but time will tell.

>>106560011 (OP)
what is gaussian splats?

>>106560166
Ask your mom

>>106560011 (OP)
It's not clear to me how long it takes to decode. And how much CPU.

I remember some anon making a top down game with gaussian splats either on our /gedg/ or /agdg/ a few months ago. Thing looked like google maps from the screenshot I remember.

>>106560011 (OP)
I dont buy it, it seems useful for single photorealistic scenes with either no object motion or a simple predetermined object motion. Like 3D dioramas of real world objects and locations. But you couldn't render a 3D object in a new environment that it wasnt photographed in using this technique, or an object moving or animating in some way that it wasnt captured in, or anything stylized. It seems like a great optimization for 3D photography and VR, but I dont see how it can be used to replace the current popular 3D rendering methods which give a lot more freedom

>>106560011 (OP)
it's only for static scenes and objects. animating this stuff with behavioral systems is impossible

remember the Unlimited Detail engine from like 15 years ago

>>106560403
I'm an orphan.

>>106560666
I 'member, I showed it to all my friends back then.
Also nice trips.

>>106560671
Oh...

>>106560011 (OP)
isn't this from that video of dr. carol something

>>106560666
>>106562343

>>106560011 (OP)
It won't make things faster. The software industry will just lay off even more people responsible for optimization and the CEO will be able to buy a little large yacht.

>>106560054
glad you attached an image to your post showing all the great things it does.

splats are pretty cool, I don't think efficiency is the big point but photorealism, at least for photo-derived materials. It's basically nerf but performant enough for real time.
Though probably it will just get leapfrogged by diffusion world models or something sadly

>>106563603
I don't need proof when I have hyperbole to back up my statements.

>>106560011 (OP)
>>106560560
>>106563558
>>106563603
This retard made a thread and forgot to attach the video he is talking about:
https://www.youtube.com/watch?v=_WjU5d26Cc4

>>106563796
gotta love this guy
https://streamable.com/wwc8gc

we used to have threads on this like 10+ years ago

https://github.com/tux3/Evolve

>>106560011 (OP)
What Andy giveth, Bill taketh away.

>>106560438
making splats is the computationally intense part, displaying splats can be done on very low-power hardware. Gaussian splats are literally just colorful blurred particles in 3D space. Animation with 3D splats is possible by reconstructing multiple frames from a video, taken by multiple angles.
Of couse you can display splats in games, but they can't receive any lighting, apart from being brighter or darker overall.

>>106560029
we are going to have AGI by then
ASI 2033
benchods will doubt

>>106560011 (OP)
In 3-4 years there will be no operating systems or apps. It will all be AI video generated on the spot simulating being an OS. Not even internet will be there, just AI agent that will scour the internet for you and then generate something similar.

This innovation is like 5-10 years too late.

It's a rendering technique that's much more efficient than before? So we basically get more performance with less power?

I implemented a gaussian splat renderer for work. They look cool but have visual artifacts even if implemented properly, use a ton of ram, don't support animation. It's better for a usecase for like scanning a 3D model and then letting people interact with it in a museum setting. Not real time rendering. Or letting the user point their smartphone at an object and convert it to 3D

>>106560438
You can do it in real time on modern gpus / igpus

>>106566087
You get more realistic 3D models, but you'll start to notice artifacts if the camera gets too close. They're better at capturing real data into a 3D form you can examine. Traditional polygonal data remains superior for most games.

>>106560011 (OP)
>Within like 3-4 years virtually every single real-time simulation or video game is going to be like 2x more efficient due to this tech
absolute fucking horseshit vomited up by a retard. it's like you retard fucking faggots forgot about the hype of voxels and it's still exceedingly rare to see games utilizing them for everything.
>>106563796
>it's just a compression algo
embarrassing. this board is a fucking joke.

>>106560166
A fancy name for calculating 3D models from a series of 2D photos. Google Maps satellite view has been using it for a decade now in bigger cities.

>>106566087
>It's a rendering technique that's much more efficient than before?
it's not more or less efficient, it's just yet another image encoding format with very good compression. it will allow your 3d graphics to have far greater detail on textures but a fraction of the size. encoding such images is insanely processor intensive. decoding them is not.

>>106566209
>I implemented a gaussian splat renderer for work. They look cool but have visual artifacts even if implemented properly, use a ton of ram, don't support animation. It's better for a usecase for like scanning a 3D model and then letting people interact with it in a museum setting. Not real time rendering. Or letting the user point their smartphone at an object and convert it to 3D

Well that's disappointing to read in a thread on /v/. It's always the same with these types of threads hyping stuff.

>>106566235
> expecting more from spamming retards on this board
are you unwell? this board is moron world headquarters.

>>106566209
>You can do it in real time on modern gpus / igpus
Sure but if it's competing against JPG it sounds kinda retarded. You can decode JPG on a toaster

>>106566383
It's not retarded and has actual applications. For example a realtor can walk through a house with an iphone, record every angle of the house, then use structure from motion to generate a splat file that can be streamed to a user so they can do a virtual tour of the house on their phone. A splat is essentially a sphere that is stretched with a specific color, when you sort them correctly and render it you can view the model from any angle. It's not possible to do this with just jpegs unless you manually create a 3D model and then uv map a million pictures of the house to the various faces. And it's still going to look way worse. Splat rendering can run at 60fps+ on a smartphone.

>>106566209
what kind of work do you do and do you have any career advice

>>106566946
graphics programming on a proprietary engine, not game dev but adjacent.

career advice:
> make a visual portfolio
> learn Vulkan and have a basic rendering engine on your github. Metal and D3D12 are nice to have but if you know at least 1 low level rendering api you can generally pick up the other two quickly
> know rendering techniques (clustered, forward+, deferred etc)
> know how compute shaders work
> be able to describe the entire graphics pipeline
> know what renderdoc is, know what tracy (gpu profiler)
> accept that these jobs are niche and its hard to get into the industry
> know C/C++ well, most rendering engines are using this on the backend.
Couldn't get a gpu job until my third one in my career. Best of luck.

>>106560011 (OP)
>believing youtube clickbait
>from a dying AI hype channel no less
Shame on you /g/

>>106567085
Do you know CUDA? would it be more lucrative for you to work on CUDA programming instead of graphics?

>>106567433
no but I work with compute shaders in vulkan intentionally and not CUDA so that our compute solutions are portable.

>would it be more lucrative
no idea, but writing code that ONLY works on one gpu vendor is imo stupid so I wouldn't do it if I had agency over the project. Which I do.

You can't recreate an accurate image from a low quality input. This is literally why data can only be compressed so far.

>>106560011 (OP)
Meh, more trash tier fake pixels.

No.
Gaussian splats are pretty good but they fall off hard when you look at different angles or positions.

>>106567975
How about I Gaussian splat my cum across your face huh?

why does my modest $900 32GB VRAM 256GB RAM PC desktop build need this again

>>106568086
Because Sam Altman is trying really, really hard to not get assassinated by his investors.

>>106568096
you're 100% right. I keep forgetting the economy runs on this dumb shit

>>106567975
This is just lack of captured data
Has nothing to do with GSP

It's like going behind or inside a building in a game where it's not modeled and going polygons suck

>>106567975
This is a new approach that is much better, which is the point.

>vector images are much smaller than regular images
>dude, what if we did 3D with them?

it's really that simple?

>>106568920
svgchads I kneel

>>106563796
>even in a horribly compressed jewtube video you can see details just plain erased from the end result
Another fucking useless piece of shit that would only impress mouthbreathing retard subhumans (aka the average /g/ poster).

Apparently the new cool thing is to use triangles instead of splats

>>106560029
Yeah. Even if the tech becomes perfect tomorrow, games take like 6-7 years to develop now, so it would take at least that long, likely an extra year for it to be implemented into game engines.

>>106569078
Yeah that was a bad example to include.
If you are not detecting high gradients like this in image geometry and absolutely erasing it from existence, your system has some big flaws.
Why these got deleted and not the small dark dots, or the long strips of lightness, that's the real question.
It has potential but until they get the accuracy actually high for details like this, it's not viable.
Even lossy jpg is good for document storage if you use a moderate compression level, imagine if some of the letters simply got chunks deleted from them because the lighting was slightly off, or the camera was angled a certain way and the font was just a little too small to be resolved properly.
Gaussians can be used to great effect, and they've been made way more useful over the last year, but they've still got some growing up to do.

>>106560666
You mean the thing that worked and all you retards got proved wrong?
You mean that thing that had entire companies taking it on and still do now?
Who are legit one of the leaders in cloud point streaming systems for holography, laser scanning and others?
Those guys? Oh yeah.
I bet you were even nodding in agreement when that fat retard Notch was attacking them as if that fuckwit knew anything about optimized code.

Protip retard, the same concept is used in UE5, just to a way better degree because it was made by graphics experts and not someone who had a hateboner for the graphics industry.
Now, the rest of UE5, that's a whole different story, but that part of UE5 is at least solid.

>>106563830
Holy shit, someone remembers!