All About Gaussians

Talk · Michael Rubloff · 3DISE 2026 · 5 May 2026 · 16:50 - 17:10 · 15 min

Chapters

What the talk is about

Michael Rubloff, founder of RadianceFields.com, gives a short history of Gaussian splatting as one way to render a radiance field and then focuses on dynamic Gaussian splatting (4DGS), essentially lifelike 3D video. The talk walks through real productions already using it, from Superman and an A$AP Rocky music video to NBA broadcasts, Apple Vision Pro personas and creators on YouTube, and where Michael expects it to go next.

Key ideas

  • Gaussian splatting was published in 2023, won Best Paper at SIGGRAPH 2023 and, according to Michael, now sees several new research papers a day.
  • Dynamic Gaussian splatting reconstructs a moving scene from multi-view camera footage, so viewers can move around inside what looks like a normal video.
  • According to Michael, the Framestore team on Superman gained creative flexibility by capturing the actors first and deciding later how to frame the scene.
  • Sports are an early use: NBA on Prime already uses Gaussian splatting to break down game highlights, and teams can replay plays in augmented reality.
  • Michael expects concerts, music videos, social media and, later this decade, feature films in dynamic 3D, along with cameras built specifically for it and native 3D content creators.
“I actually do believe that dynamic Gaussian splatting is going to really emotionally resonate with people”
— Michael Rubloff ·
“I think that we're gonna start to have people emerge who are native 3D content creators as well pretty shortly.”
— Michael Rubloff ·
“anything that currently is being represented in 2D, uh, will be represented in 3D.”
— Michael Rubloff ·

Questions answered

What is dynamic Gaussian splatting or 4DGS?

Michael Rubloff described it as Gaussian splatting extended to video: a moving scene reconstructed from multi-view images, so you can move around inside it like 3D video.

Has dynamic Gaussian splatting been used in real productions?

Yes. According to Michael, it was used for Superman, the A$AP Rocky music video Helicopter, NBA on Prime highlight breakdowns and Apple Vision Pro personas.

Can dynamic Gaussian splatting be used in architecture or engineering projects?

Michael suggested arranging repeated static captures of a construction site into a time-lapse flip book, or placing a dynamic capture of a person inside a static site reconstruction.

Full transcriptabout 14 min readShowHide

Transcribed automatically (ElevenLabs Scribe) and lightly edited. It may contain small errors. Click a time to play the video from there.

Host: Welcome.

Michael Rubloff: Thank you. Thank you. All right. Well, thank you so much everyone for coming to listen to my talk today. I'm really excited to talk a little bit more about some of the things that might be coming up in the world of Gaussian splatting.

Michael Rubloff: So let me, without further ado, just start going through it. So, um, okay, let me change the slide. There we go. Okay, great. So to kind of, before I get into the talk, uh, give a little bit of information about myself.

Michael Rubloff: So my name is Michael Rubloff. I run a website called radiancefields.com that I started just at the very end of 2022. At that time, it was pre-Gaussian splatting. It was only NeRF.

Michael Rubloff: And what has happened since then is that the technology has rapidly spiraled out of, uh, control in terms of its capabilities. So to date, I've posted, uh, or published a little bit over 900 articles about both NeRFs and Gaussian splatting, and have also tried to service as much as I can to help the spread of lifelike 3D become, uh, everywhere. And so I also, to that same end, uh, host a job board where people can find the latest jobs in Gaussian splatting and have also created a daily paper tracker as well. And then I also provide consulting services and quality control for businesses that are looking to get into more of the Gaussian splatting world, and also do commission scan work too. So with this all in mind, let's talk a little bit about the world of imaging.

Michael Rubloff: And so I always really love this quote, which comes from, uh, Leonardo da Vinci, where he said, "The first intention of a painter is to make a flat surface display a body as if it was modeled and separated from this plane, and he who surpasses others in this skill deserves the most praise." And so we've seen there's been a lot of work to try and move towards representing the world in a, a 3D capacity. And, you know, post-painting, we move to the world of modern imaging, uh, which we would first have with say, like black and white and film, and now we've moved to computational and digital photography, as from this example of my dog on the right. But as it turns out, you know, we also have the way to represent the world through video too, and we can see this example here is an actual video of a street in London. But as this turns out, this is actually not a normal video.

Michael Rubloff: This is actually a Gaussian splat. And we can see some of the fidelity of which we're able to achieve using Gaussian splatting and basically, uh, Greater Turnip put this into Unity with another plugin, and so he's able to simulate walking through. So that's why it really just looks like we're looking at a normal 2D video here. But again, this is complete static 3D. And to give a quick overview about Gaussian splatting, Gaussian splatting is one way to represent a radiance field, otherwise known as a very lifelike 3D representation.

Michael Rubloff: And it was published in late April 2023, meaning it just hit its three-year birthday just, uh, about like a week or so ago. And then it was recognized at SIGGRAPH 2023, winning Best Paper. And since then, there's been a lot of significant advancements from pretty much every single component of Gaussian splatting, ranging from compression to rendering rates to fidelity, and tooling too has also radically increased in its capability. And, uh, I always love this, uh, this last little stat here is that it was the 19th most cited paper of 2023. Number 17 was Google's Gemini, to give context as to how much interest there is in Gaussian splatting.

Michael Rubloff: And last year there were, uh, 4.63 papers published, uh, every ... or pub- published a day, including both weekends and holidays. So it's starting to really pick up in terms of, uh, its adoption in the research world. But we can just, you know, take a look again.

Michael Rubloff: The full name of Gaussian splatting is 3D Gaussian Splatting for Real-Time Radiance Field Rendering. It is a way to render a radiance field. And as it turns out, there's actually quite a lot of different ways to do this. We really have a strong understanding now of how to reconstruct the world around us in extraordinarily lifelike 3D. And while Gaussian splatting is a very hot topic and, and term, it is but one way to reconstruct the world around us.

Michael Rubloff: So as it-- as mentioned earlier, we also have the capability of representing the world in video form. And we can see here, you know, this is Henry from Infinite Realities, uh, making a bunch of different, uh, uh, I guess, faces. But as it turns out, this is actually not a 2D video either. This is actually dynamic 3D reconstruction, meaning that we can actually move around the space. It essentially is akin to 3D video.

Michael Rubloff: And for that, that's actually what I'm gonna focus this talk on today is one of the emerging parts of Gaussian splatting, which is not the static one, which you've seen a lot of examples of today, but more moving into the world of volumetric video or dynamic Gaussian splatting, which is also referred to as 4DGS. And it works in a similar way to, uh, static Gaussian splatting, where we are gonna pull a bunch of multi-view images. Here's a capture of me actually, where, uh, these are all the different setups here. And from that, we're able to reconstruct it just similar to, uh, how static Gaussian splatting works as well. But now we can see instead of being in a static moment in time, we're able to actually expand it to be a very lifelike video.

Michael Rubloff: And it has really grown in its capability. So this was taken back in 2024 and, um, it's really nice that, uh, of the starting point, but the fidelity actually has continued to improve. So this was two weeks ago at NAB, where we can see now the fidelity is continuing to get stronger and stronger to the point where it truly just looks like we are watching 2D video. And I'm very, uh, positive that this, this capability is going to increase in capabilities. And so this was in a very small booth, and so you really can set this up in a small footprint too.

Michael Rubloff: You know, we're ... You can see the box that I'm standing in. It is a very tiny area. But one of the things I wanna actually talk about here is where this is heading. And I, I actually do believe that dynamic Gaussian splatting is going to really emotionally resonate with people and start to make people really request and want to experience lifelike 3D in their daily lives.

Michael Rubloff: And some of the ways that I think this is going to happen over the second half of this year and going into next year is going to be we're gonna see more examples from concerts are going to, uh, start to be released, I think, in dynamic Gaussian splatting. Uh, we're gonna see more music videos. Uh, there's a couple examples in, uh, a couple slides from now, taking a look on the music video side. Uh, and I think that, you know, towards the, the back half or later part of this decade, we're gonna start to see larger projects going to the scale of feature films in dynamic 3D as well. But the last one too I'm also really excited about is the YouTube and social media applications of dynamic Gaussian splatting, where we're gonna start to see it becoming more feasible for, uh, consumers to actually represent and capture moments in time in a video capacity.

Michael Rubloff: Uh, you might notice as well I'm wearing the, the same jacket in this. Uh, and you know, just like how we, we have gotten s- uh, specialized hardware for static Gaussian splatting, I think that it will be very likely that we will start to see companies build, uh, cameras designed specifically for this application as well. So to talk about, you know, would this actually come out anywhere in the real world? And the answer would, uh, be yes. It actually has made its way already into a couple of very large, big, big budget projects.

Michael Rubloff: One of the, the first one, it was, uh, Superman, which I mentioned in the panel earlier today. And you can see these two captures or these two figures of Superman's parents. Um, none of this is 2D. This is all complete dynamic 3D reconstruction using Gaussian splatting. And when I spoke to the Framestore team behind this, uh, what they actually told me was that it gave them significantly more flexibility in their actual creative workflow because what they decided to do was they just filmed everything up front.

Michael Rubloff: They, they just went and captured the actors, and then from there, they decided how best to actually display the, the, uh, the scene in, in its final form based upon everything else. And so that was one of the first early examples of dynamic Gaussian splatting. But it was, again, this was captured nearly two or over two years ago at this point. And so things have definitely continued to move forward. The most recent example of which is, uh, the actually the music video for A$AP Rocky's "Helicopter," where pretty much every single person in this music video is dynamic Gaussian splatting as well.

Michael Rubloff: So all the things you see here of the crowds and people, all of it is dynamic 3D, and they shot it with, uh, a fifty-six RGBD camera array that was then used. And so from what I understand, there's quite a lot of footage from those shoots that will continue to hopefully make their way out into the world, and it was done in Houdini and Octane as well. Uh, another direction that I think is really interesting to see is the world of sports and dynamic 3D, where we're starting to see companies like Arcturus Sports, who started to release some of their examples, where we can create truly incredible camera angles like this one. Uh, and we're able to actually really get very close to the moment in time for which something is occurring, but now being, uh, be in 3D. And we can see, we could track the ball, we can go right up into the actual, uh, location.

Michael Rubloff: And so this is all just, like, very crazy that we have the capabilities. But if you're kinda looking at this and being like, "Okay, maybe this is just like a demo, you know, when were we gonna see this out in the real world or on TV ourself?" Uh, it actually has already happened as well. So here is NBA on Prime from Amazon. We can see then the post game.

Michael Rubloff: They're actually using Gaussian splatting now to break down some of the highlights of the game. So this is from a real NBA game. There's actually quite a few clips that they've posted at this point using Gaussian splatting, and we can start to get a better sense as to where this is actually going to be used and how it's gonna be used. So in the case of sports, obviously, we have it from the consumer application where we can go and get closer to the action, but then we can also look at it from the team angle, where they're actually able to then simulate and, uh, play back the actual game onto the floor itself in, you know, say, augmented reality. And so there's a lot of very, very crazy use cases that have not really been fully leveraged yet.

Michael Rubloff: But as it turns out, there's also a few from the world of big tech where we can see actually Apple has officially confirmed that they are using Gaussian splatting to power their personas in the Vision Pro, which essentially are these very lifelike dynamic 3D avatars. Uh, this is Brad Lynch on, on the left here, and you can see it looks very lifelike. Uh, a- and they've only continued to get more and more powerful, and it would not surprise me if later this year we see Apple talk more about Gaussian splatting as well. Uh, one other note too is that the iPhone 17 Pros now do support Genlock, and we actually have begun to see people create these camera rigs using-- uh, to be used with Gaussian splatting. Uh, there's a company called Radiant Images who has a rig called Meridian that is also, uh, doing this.

Michael Rubloff: And on the right, there's also Meta's Codec Avatar program, which is also similarly, uh, able to create relightable and very lifelike 3D avatars using Gaussian splatting as well. Uh, some other recent examples are from Netflix's Eyeline Studios have a method called detail-enhanced Gaussian splatting, where they have a very expensive camera studio where they are able to essentially also film in 3D and capture and be able to composite and reinsert people into a location to the point where you really can't tell they're not physically there. And finally, NVIDIA also has the IMU Lab, which, uh, open sourced-- is, uh, one of the only open-sourced dynamic Gaussian splatting repos out there called Queen, and that is open sourced, and I believe pretty permissibly licensed to use as well. But some of these other fun examples is, uh, on the YouTube side, and this is where some of the content creators are, are starting to move into really being able to actually create content in dynamic 3D. So this is a creator named Tim, otherwise known as MediaStorm, uh, who is a very, very large tech influencer in, uh, China.

Michael Rubloff: And we can see the fidelity in this is just amazing, uh, that you're able to move up very close to him. And yeah, we can see that the fidelity really holds up. And so I'm hopeful that we're going to see just like how we've had, uh, creators, you know, in, in radio and like, say, black and white, uh, film. I, I think that we're gonna start to have people emerge who are native 3D content creators as well pretty shortly. Uh, a second example of this is actually Ren from Corridor Digital.

Michael Rubloff: Uh, so actually the last one and this one too are both being filmed and reconstructed from a company called 4DV AI. And we can see too, so Ren is actually not actually in this room, and he's been reinserted into it. But we can see there's a lot of creative flexibility that you get when you actually are filming everything in 3D. And we can see that as well. We can slow down time.

Michael Rubloff: We're able to represent a really, a, a longer moment in time and then reinsert into different locations like that, and we can do it in a pretty seamless way already to this point. And so all these are splats and, um, so again, I really would not be surprised if we start to see just fully native 3D content creators in just a little bit. So with all of this in mind, um, everything is moving very quickly in the world of Gaussian splatting, both on the static side and the dynamic side. And this is what I spend my entire life on, is essentially cataloging this technology and staying up to date with it. So, um, a-as I mentioned at the beginning of the talk, I post pretty much daily on the website, and I've been trying to make sure that I don't miss out on anything.

Michael Rubloff: So if you're interested in learning about where Gaussian splatting is headed and don't wanna fall behind, I highly would recommend, uh, consider following me for, for this purpose. And on the same side, I'm also being-- I have also been creating a lot of content over on YouTube to sh- actually highlight some of the applications that were shown in here. So, you know, my goal is to really make sure that we are getting to a world where lifelike 3D imaging is natively how we interact with the world. And so hopefully I can, uh, build out that, that future for, for everyone. So if you're using Gaussian splatting or any lifelike 3D representation, please feel free to reach out to me.

Michael Rubloff: Uh, otherwise, that is my talk.

Host: Michael. Michael, Michael, Michael, thank you so much. We have some spare time for two questions. So the first one, actually two of the same kind. Guys, I can see what's on your mind.

Host: Let's put the, the, uh, the, the Slido up there. So wasn't there one category missing from dynamic GS that will probably be the most popular? Uh, wink, wink, there's another one talking about, um, you know, adult content, let's put it this way. What are your thoughts on that?

Michael Rubloff: Yeah. Uh, my thoughts are that anything that currently is being represented in 2D, uh, will be represented in 3D. And so I think that, yeah, any, any of these use cases will, will apply.

Host: Will, will follow.

Michael Rubloff: Mm-hmm.

Host: Um, okay. And, um, let's take on, uh, this one. Can dynamic, uh, 3D GS be integrated into architecture or engineering projects, and if so, how?

Michael Rubloff: Mm-hmm. Yeah, I think it definitely can. Um, in terms of-- There are a couple different ways that you could try and, and do this depending on how complex you wanna get with it. Um, in terms... Well, th-there's actually a few different ways.

Michael Rubloff: There, there's one in terms of, uh, dynamic in the sense of, say, if you take multiple static 3D captures, and you start to arrange them into a flip book and create a dy- almost a dynamic, uh, uh, time lapse of, let's say, a construction site or being able to peel back to do maintenance. That's one way that you could do it. Um, but if you wanted to really truly have dynamic 3D Gaussian splatting, you could, uh, take a static reconstruction of a construction site and then take a dynamic reconstruction of, let's say, a person, and then have them walk through that actual location itself too. So there, there are a couple different ways I think, uh, you could potentially look into integrating that.

Host: Michael, thank you so much, and also thank you for sharing all that content on a daily basis and inspiring us all. Let's give him a big round of applause. Michael Rubloff , thanks.

Michael Rubloff: Thank you so much.

Speaker

Michael Rubloff

Michael RubloffLinkedIn

Founder @ Radiancefields.com | NeRFs & Gaussian Splatting

Michael Rubloff is the Founder and Managing Editor of RadianceFields.com, the leading publication chronicling the rise of radiance field technologies, and the co-host of the View Dependent podcast. Michael speaks, consults, and collaborates with companies regarding the impact of radiance field…

Hear the next ones live – 3DISE 2027, 20–21 April, Prague

Buy tickets