NVIDIA's New AI: 150x Faster Virtual Worlds!
Вставка
- Опубліковано 5 лют 2025
- ❤️ Check out Lambda here and sign up for their GPU Cloud: lambdalabs.com...
📝 The paper "InstantSplat: Sparse-view SfM-free Gaussian Splatting in Seconds" is available here:
instantsplat.g...
Try it out (hopefully works): huggingface.co...
Clouds paper: arcanous98.git...
📝 My paper on simulations that look almost like reality is available for free here:
rdcu.be/cWPfD
Or this is the orig. Nature Physics link with clickable citations:
www.nature.com...
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Benji Rabhan, B Shang, Christian Ahlin, Gordon Child, John Le, Juan Benet, Kyle Davis, Loyal Alchemist, Lukas Biewald, Michael Tedder, Owen Skarpness, Richard Sundvall, Steef, Taras Bobrovytsky, Thomas Krcmar, Tybie Fitzhugh, Ueli Gallizzi
If you wish to appear here or pick up other perks, click here: / twominutepapers
My research: cg.tuwien.ac.a...
X/Twitter: / twominutepapers
Thumbnail design: Felícia Zsolnai-Fehér - felicia.hu
#nvidia
Oh hell yeah
The "artifact" at 2:49 is actually the watermark from the original video made by Open AI Sora.
I was thinking the same thing
Really
That was pretty obvious for anyone who followed the Sora story or has watched some Sora videos, and it makes me wonder how or why he would make such a rudimentary mistake, or what exactly was going on there. Some of the 3D worlds in this video seem to be generated directly from old Sora videos (taking 3 frames and generating a 3d environment from them?), and he should have talked about that to clear up exactly what he was trying to show.
This gonna be huge for VR viewings, for normal games it's probably still need a ton of works for devs
XR - for the instant collision maps of the environment
Should be great for quick mock ups to get a feel for a certain scene before committing to more detailed and cleaned up models. You can probably also take a few screenshots of other games/environments so you can 'borrow' a certain look, adjust it and it's a new scene/object.
This makes sense why I got an email from Adobe saying “Introducing Unmetered Access to Substance 3D Assets.” haha😅
As someone working with 3d for 20 years all of these new techniques coming out are true magic. Photography of the future will be truly interactive. You will be able to step into a picture and interact with everything
What a time to be alive!
What a Time to be AI😂
From instant NERFs to instant Splats in less than 3 years. What a time to be alive!
The version I want to see is handing it a video and having it reconstruct the world and everything in it.
so happy to see gaussian splatting paper, was really amazed by the GS and its so cool to see it progress
41 ai videos in the last 6 months, i'm tired boss.
I hope this tech someday lets us turn 2d movies into 3d for VR, that would be amazing!
It feels like it was just yesterday we were talking about photogrammetry and now here we are getting our minds blown by NeRFs and Gaussian Splatting and now they're like take 3 pictures and I'll clone you 😂😂😂😂
Reminds me of George Carlin:
"Soon all you'll need to get to college is a pen. You got a pen? Get the fuck in there, it's physics."
Soon all you'll need to do anything you want is a computer with AI.
When you see a 2-minute papers video uploaded 1-minute ago, jump-in, you've got 60 seconds left!
Gotta get it while it's fresh!
0:39 "A BLURRY MESS" ... just like DLSS
Wow this is truly impressive how fast these algorithims are going, Nerfs are really cool for creating real places to visit virtually
TMP is such an awesome channel! i now know how to say your name out loud :D
Imagine Google maps street view with this tech.
Me: make me an oscar award winning movie.
AI: do you want me to help you with the winning award speech as well?
DID not you forget to do videos on the Chinese AI?
Alibaba has released a new version that combines chat bot with image generation.
One big step in 3D rendering and one small step toward the Holodeck.
I tested InstantSplat last autumn for my projects, and while the individual splats can look nice (and the initial sparse model looks clean compared to Colmap, when just eyeballing it), the geometry/calibration is actually very inaccurate, and you can't easily e.g. combine overlapping splats to create larger models.
I really don't see much use for that technology right now.
The reason being that the produced 3d-models are a huge mess and pretty much impossible to work with. They are a solution looking for a problem to solve.
What the people in 3d graphics REALLY want is a software that can deconstruct the photographed real world objects into simple base shapes, of precize and clean topology and low poly enough to be usefull within the 3d-graphics workflow.
the demo isnt working :'( this looks stunning
Can confirm. Failed on hugging face
It's cool to see photogrammetry make such leaps due to AI. Same with selection tools for image editing.
Ezt lehetne pontfelhőknél alkalmazni? Pontfelhőből 3d-s modellek létrehozására?
"There's no end to the possibilities!"
All this digital stuff is nice, but oh, i yearn for Bio-med breakthroughs
So true, where is zerglike ai? Fungal ai is the closest we have.
Nvidea unveiling this so their stock price goes back up
They unveiled this a month ago. The code and paper have been on their Github since Dec. 30.
YEP LMFAO CHINA FKD UP EVERYTHING
Like magic! This is what I've hoped for! Now just need to get it in a smooth interface that works right off of phones and on to AR glasses!
This is not 3D modeling. Please stop misleading people.
lol
like the old school snl references in your edits
tunces the driving cat
I can't wait until I can take two photos of an object, press a button or two and print it on the 3d printer.
If you want to make a doom level, use screenshots from videos of Jensen Huang's kitchen.
michael bay just pop the champagne bottle somewhere! what a time to be alive!!!
It's been a dozen times I see here that Nvidia AI is now 150x times faster. So now it should be billions of times faster than 2 or 3 years ago 😂
micheal bay is actually pretty fervently anti-ai but i do appreciate the application possibilities here for things like mapping dangerous areas out or exploring the terrain of planets
So a full 360 of images will result in a perfect model of the scene? Photogramertry got a massive upgrade!
We will be able to live in our dreams. We will create an image of the world we have dreamed of by describing it, and then it will materialize before our eyes in VR
I'm sorry, but I'm going to unsub. Your early day reporting was so much fun and for a long time you were the primary channel to report advancements in (ai) technology. But nowadays this channel is late and undetailed in reporting. There are so many channels now that show ai news as soon as it comes out, not two papers down the line. Also, this has become an Nvidia fanboy channel.
The irony of this video coming out after Deep Seek caused Nvidia stock to crash
Where's the irony, Alanis?
@@notthatrkelly I don't think they know what this word means.
Deep Seek running on NVIDIA hardware. People need to change perspective. Ai just got 10 times more powerful.
This is unbelievable. “Paradigm shift” is absolutely correct
Well.. if we can do it in real time.. then holographic television is finally here :) That is a big plus.. so let´s shave down those 12 seconds to 1/25th..
This is perfect to generate the cities street views Google and Apple Maps apps. I hope they use it.
Can I run it for 90 seconds and get one without artifacts? Messy 3D models are still a pain to work with.
If you played a video game in an environment scanned from a real world place,
you would quickly realize the importance of level design.
they could build realworld sets like movies do that are based on good level design then scan them in for the best of both worlds
Has there been an update, as it seems that this is from 7 months ago? Im wanting to test it out so am downloading now.
Why did you take the Deepseek video down?
I really feel at the edge of technology when I tinkered around with something on my PC and 2 minute Papers makes a review of this _afterwards_. :D
A splat is not a 3d model! You should talk voxels if you really talk about 3d modelling. Only low resolution 3d models can be retrieved from a splat, which looks nothing as great as the actual splat.
kiri is able to pull models from splats that are decent
@@zethhamblin1587 Trellis is not too bad either, and can run locally. I'm just hoping for more resolution.
What's the news when it comes to Deepseek? 😮😮😮😂 It's an explosion professor I def. Tough that would be todays news :)
1:40 has definitely more than 2 views by now
/s
In two minute papers, we run things in or kitchen on DOOM, not run things DOOM on things in our kitchen!
ALL HAIL THE A.I. OVERLORDS !!! WHAT A TIME TO BE ALIVE INDEED !!!
Waiting for ultra-fast image generation, sCM or something. Any news?
wow AI and a big number, must consoom video
Won't be long before it'll be possible to accurately model almost any location on earth with one or two photos.
Is it posible to make with 360 images?
Now we'll get a decent Wolfenstein!
I have just given up chasing and learning deep learning and AI at this point. Its impossible to keep up.
Making indy games is going to be so much fun going forward, no great idea will be wasted
Actually this channel was much much better in the past. Now it's just clickbaits, pop trending AI things.
Mr Károly is not the same interesting presenter as it was.
now all indy games will look the same
Could be good for turning regular video into stereo 3D video.
Where is the bump back up in NVDA stock?
Two Minute marqueting.
I want to see the 3 images used for each video
What a Switch Two Be Alive !
i couldn't get the hugging face instasplat to complete a model, so i downloaded and opened the santorini point cloud file (.ply) in blender (using import). the point cloud opens, but so far, not been able to show the vertex colours. however, this is clearly progress, though having a full mesh would be ideal (which i know other apps/papers can do).
InstaSplat uses Gaussian Splatting which is based on ellipsoids, gaussian distributions and a view-dependent technique for the coloring, you won't get a mesh here like in traditional workflows. So you'd need to switch to e.g. a Blender plugin for viewing/editing the Gaussian Splats.
@@schnaader i wasn't expecting a mesh, but i had hoped to see vertex colours. however, as you suggest, i will need an addon for that.
So that scene in Blade Runner was simply AI....hmmm.
Want to run it locally… just buy our RTX 9999 for $19999 per day
Is this a switch 2 ad?
Actually this channel was much much better in the past. Now it's just clickbaits, pop trending AI things.
Mr Károly is not the same interesting presenter as it was.
Of course he will not like this comment, just hope he will acknowledge it.
I do miss when this channel had things outside of AI. You're a brilliant mind Károly! Please diversify your content more.
Did he ever do something outside of computer graphics and ai?
@monsieurLDN graphics specifically is what I'm interested in. He used to cover tons of simulation advancements like a baking bread sim, cloth simulations, water simulations and it was way more interesting from someone involved in the space.
Before the ChatGPT and mid journey explosion he used to cover tons of other topics
fake dosent work it needs a gpu quota so high so much for open source "You have exceeded your GPU quota (150s requested vs. 31s left). Try again in 23:05:01"
I can't just not click on a thumbnail for Schmitt's Gay.
can't AI already do videos from 1 picture though
I did a scale improvement for an old picture in an Ai webstie and it got a bit better, I remarked the Ai played with some parts of it when it is blurry and it changed to other thing like a steel bar standing in the road to a human being walking around ...It just copies from somewhere to the picture as it does understand it ! nothing magical, it is just science and doom the human job and his/her salary !
I wonder how much of this speedup is fully algorithmic vs. just faster compute hardware?
RTX 50XX might run it blazing fast, but doubtful anywhere near that fast on an RTX 30XX, if at all. (Some RTXs are more RTX than others it seems).
🖖😎👍
The main difference between 50XX and 40XX are a couple of newly designed chips. that why the 5080 can compete with the 4090 in gaming.
@@nox5555 I suspect that much more emphasis is being placed on the AI component than the game render architecture these days. (GAI domination of the world being a slightly higher priority than making zombies and waifu and look more photorealistic.)
I'm gonna create a game.
Well, it is gaussian splats.. not 3D modelling? right.. It is a glorified pointcloud, with very limited range of visibility and rather weird artefacts just outside the range of the original angles.. If it were 3D modelling, I´d be very impressed.
You're looking at the idea of things. Not the things in-of themselves.
Can we stop calling things that are not intelligent AI, advanced algorithms I grant you, but intelligence? No.
Splatting is not modelling, you can't really edit it and it's useless on CGI / Game dev pipelines
hmmm... buying some Nvidia Shares now when they're cheap again ??? 🤔
Yes ? No ? ...hmmm...
Hell yeah! now things are going to get 150x cheaper for us, the consumer!
Right?
I need Nvidia's stock to go 150x. lol
3:29 Nice joke, lol
Nvidia CANT Fix the way this guy talks...
looks a bit like DLSS
Enhance.......enhance.......
balls
Very cool, still looks like shit tho. Also the fact that you're not, in fact, looking at a 3D reconstruction of the scene but rather a 3D reconstruction of whatever the AI thinks the scene looked like. I'm sure this has it's applications but I wish NVIDIA would go eat shit with all their hyping of model-generated content.
Do deep seek!!!!
This is not 3D modelling, liar. This is useless gaussian splat. Classic photogrammetry can do much better.
You're an idiot. Let's see you do "classic photogrammetry" with three images as the input. You can use gaussian splats in Blender, Unity, Unreal Engine and there are other programs to create textured meshes from them.
with 3 photos? just because splats aren't useful now doesn't mean that they won't be useful in the future... I'll bet we'll be able to pull 3d models from them soon or just be able to work with spats the same way we do with 3d models after some tools are developed
@@zethhamblin1587 nah, it was just an artifact of an algorithm. an extension of a point cloud but no final merit. we only see with two eyes and make much better mental models than mere smeared dots. reconstructing the actual geometry is far more valuable. The (intelligent) world is not a collection of disjointed grains.
Dudes think this is badass
I'm loving it. I might buy a drone, get A LOT of good pictures/video of buildings, to then put in city builder game/simulator :D
If it can easily be converted and exported 3d models that is
there is software to turn this into mesh. but the real game changer will come in 10 years, when this can be integrated into a city builder to make everything look realistic instead of crazy roads and mountains.
Exponential progress happening right before our eyes. It will be much crazier this time around
wrong video for me
not interested
6 minute papers
This technique has basically reached the fidelity of hand crafted 3D models from a talented artist.
I'd say it's still very, VERY far from it
omaga