Panel: Gaussian Splatting in Tech and Creative Industry

Panel discussion · Michael Rubloff, Mindy Li, Yoshiharu (Josh) Sato and Tomas Barnas · 3DISE 2025 · 25 April 2025 · 11:50 - 12:30 · 32 min

Chapters

What the talk is about

Moderated by Michal Gula, Michael Rubloff (radiancefields.com), Mindy Li (XGRIDS), Josh Sato (Waldek Technologies) and Tomas Barnas (Overhead4D) explain what Gaussian splatting is, how lidar makes splats measurable, how splats are shared on the web and shown in VR, and how to get started. The panel also covers capture tips, current limitations, the industries adopting the technology and why photogrammetry and Gaussian splatting are converging.

Key ideas

  • According to Michael Rubloff, Gaussian splatting reconstructs lifelike 3D from ordinary photos or video in real time and models view-dependent effects such as reflections and glass that are tricky for photogrammetry.
  • Mindy Li explains that XGRIDS aligns lidar point clouds with Gaussian splats, so users can take centimeter-level measurements directly in the splat.
  • Josh Sato says file size is the main obstacle to sharing splats on the web, and that splats run on almost any device because they use plain GPU rasterization without a neural network.
  • Sharp input matters more than quantity: Michael Rubloff advises removing motion-blurred frames, and Tomas Barnas warns that too many images multiply misalignments.
  • Tomas Barnas expects photogrammetry and Gaussian splatting to merge into one technology with two different outputs, as photogrammetry software adds splat support.
“it's not uncommon at all anymore for there to be five plus papers every single day that releases just specifically within Gaussian splatting”
— Michael Rubloff ·
“So there is no, uh, costly neural network or AI. It's just traditional, uh, rasterization in the, uh, original, uh, GPU pipeline.”
— Yoshiharu (Josh) Sato ·
“For me, it's, it's, it's becoming one technology which has just two different results.”
— Tomas Barnas ·

Questions answered

What is the best way to get started with Gaussian splatting?

Michael Rubloff recommends starting small with a well-curated set of sharp images, for example of a single object, and processing them in tools such as Postshot, Nerfstudio or the Scaniverse phone app; Josh Sato also names Scaniverse as the best free entry point.

Can you take accurate measurements in a Gaussian splat?

According to Mindy Li, XGRIDS aligns lidar point clouds with the Gaussian splats, which allows measurements at centimeter-level accuracy directly in the splat.

Can Gaussian splats be viewed in VR?

Yes. Tomas Barnas said splats can run natively on headsets or be streamed from a PC, and Mindy Li reported that Apple Vision Pro gave the best results, with Meta Quest 3 and Pico working via streaming.

Full transcriptabout 26 min readShowHide

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

Host: We're going to now be talking about Gaussian splatting. And, uh, ladies and gentlemen, take out your phones. We have a question for you. We wanna find out how familiar you are with Gaussian splatting.

Moderator: Okay.

Host: So you use it, you've explored it, you have already heard it, or you have no idea what is it. And our upcoming panel is going to discuss its use in the tech and creative industry, but let's just paint that picture. What kind of experience you have with the Gaussian splatting? Okay, let's take a look. Okay, 37%, 38 heard of it.

Host: 28, roughly one-third, explored with it. One quarter, more than one quarter is already using it. No idea what it is. Perfect. Please keep on sending those poll votes, but also keep on sending those questions because our next panelists are going to take them and address them.

Host: And having said that, let me welcome on stage the moderator of the upcoming panel discussion, Michal Gula. Let's give it up for him, ladies and gentlemen. Michal, welcome.

Moderator: Thank you.

Host: The stage is yours. So is this iPad.

Moderator: Thank you so much.

Host: Take it away.

Moderator: Should I introduce the rest of them?

Host: Absolutely.

Moderator: Okay. Uh, welcome Michael Rubloff from radiancefields.com.

Michael Rubloff: Whoo-hoo.

Moderator: He's the researcher. He certainly is able to explain what does it mean, Gaussian splattings. Uh, Mindy Li from XGRIDS. You can meet them, uh, outside in the booth.

Michael Rubloff: Whoo.

Moderator: Welcome. Please. Uh, Josh Sato from, uh, Gauzilla Pro. He will tell you how to actually share Gaussian splattings. Welcome.

Moderator: And Tomas Barnas, you already know him. Uh, the- ... photogrammetry VR/AR professional with the real applications of Gaussian splattings in commercial world, so.

Panel: Thank you.

Moderator: Thank you, guys. Take your seats. Okay. I, I think I have not butchered your names or your companies, so I hope the, the introduction went well. Uh, there was only like a 10%, uh, is there, uh, only like a 10% of, of people that have no idea of what Gaussian splatting is.

Moderator: I think it's very important to, to start with, uh, start with that topic because most of the, uh, people here, they already, um, touched photogrammetry. So maybe it would be good just to s- like very briefly compare it and, and just to explain from your own, uh, views, what do you think the Gaussian splatting is and how it can help the, the industry? Maybe you can start, Michael.

Michael Rubloff: Yeah, definitely. So Gaussian splatting, the full paper name is 3D Gaussian Splatting for Real-Time Radiance Field Rendering. It is a way to render a radiance field in real time. So if like during the co-course of this talk, I just referred to radiance fields, I'm including Gaussian splatting in-inside of that. And basically, what Gaussian splatting allows you to do is you are able to take just a normal series of 2D images or normal 2D video, and you're able to then reconstruct very lifelike 3D from that.

Michael Rubloff: And so radiance fields are able to, uh, show what are called view-dependent effects, which basically, you know, if you were walking around, uh, last night after the rain and taking a look at like the, the stone and, uh, the s- the stone ground, you could see these, the way that light bounces and reflects, and we have this very natural inclination of what that should look like in the real world. And Gaussian splatting and larger radiance field representations are able to model these, these effects, and so we can see glass and, uh, light and, you know, shadows and things that might be, uh, traditionally tricky with, with say like photogrammetry. And it's become a extremely exciting research topic since, uh, the introduction of NeRF in 2020, and there's been a lot of additional radiance field representations. But definitely Gaussian splatting has really captured, uh, the, the industry's attention.

Moderator: Perfect. Uh, is it safe to say, uh, you're a researcher of the 3D Gaussian splattings?

Michael Rubloff: Uh, I, I would say that I try to keep up to date the best that I can with the state of the, the research. Um, it's not uncommon at all anymore for there to be five plus papers every single day that releases just specifically within Gaussian splatting, but that also does not include NeRF and some of the other methods that are in the, the, the radiance field world.

Moderator: Yeah.

Michael Rubloff: So I, I wouldn't say I'm necessarily a researcher, but I've become very friendly with the s- or state of the research.

Moderator: Thank you very much for stealing my question because I was- ... wanting to ask you, uh, how often there is a new paper because-

Michael Rubloff: Yeah

Moderator: it's, it's very often. So, so you said like every five days or so?

Michael Rubloff: Uh, five every day.

Moderator: Oh, five every day.

Michael Rubloff: For, for just Gaussian splatting. Yeah, yeah, yeah.

Moderator: Okay.

Michael Rubloff: It's, it's very active.

Moderator: So don't, don't worry about Gaussian splatting. You, you, you wouldn't keep up with that. Uh, Mindy, please, uh, your perspective will be different because you are producers of softwares and the, and the hardwares. Uh, how, how do you see and what was the idea behind... Because you, you produce SLAM scanners, right?

Moderator: So when and how the, the idea came to, to actually integrate it with the Gaussian splattings?

Mindy Li: Yeah. Firstly, XGRIDS is a company manufacture SLAM scanner, and we found that SLAM scanner is a way to, uh, quickly re- uh, capture the reality. And but i-in the past, we used LiDAR and, uh, combined different sensors together when Gaussian splatting can. We find that's the, the best way we can find at this moment to represent the reality. And we find, uh, you can use, uh, photos and videos to do the reality capture, but it's not efficiently enough, especially for those spatial modeling things that you want to, uh, scan a relatively large scale set.

Mindy Li: Uh, but if you use a SLAM scanner to do that, you can, uh, to do that reality capture in a mobile way so that you can do infinite, uh, te-technically se- uh, speaking, it's infinite scans for the whole world. So that's why we specialize in, uh, SLAM and, uh, we combine the SLAM technology together with the Gaussian splatting. And we find that LiDAR, uh, itself can help the Gaussian splat itself very well to help the photos and, uh, the poses can find their position well. So that's why the, the results can be cleaned and structured very well.

Moderator: Mm-hmm. Perfect. Uh, Josh, from, from your point of view because, uh, we, we all know that the, uh, the sharing, displaying and, and, and just working with 3D assets, uh, at all, it's, it's very challenging, uh, task. Uh, even though, uh, e-even if you want to do it, uh, online especially. So, uh, what were the challenges for you to, to be able to, to build a platform and to be able to share Gaussian splattings?

Mindy Li: Um, so firstly we-

Moderator: Sorry, it was for Josh.

Mindy Li: Sorry.

Yoshiharu (Josh) Sato: No worries.

Moderator: Sorry, Mindy.

Yoshiharu (Josh) Sato: Uh, can you hear me? Okay. Uh, so yeah, one of the, uh, main challenges, uh, of developing, uh, Gaussian splatting, uh, web-based viewer and editor is the file size of Gaussian splats. And by default, the Gaussian, Gaussian splats are so huge. Uh, it's like giga- multiple gigabytes, and it's very tough to, uh, you know, share all these, uh, you know, large, uh, they're large files o-on the internet.

Yoshiharu (Josh) Sato: So compression is one big, uh, important factor for, uh, for, you know, in a, in the very Gaussian splatting, uh, uh, web sharing. And I think, uh, XGRIDS are already doing great job at, uh, you know, compressing all these, uh, Gaussian splat data. And also there's, uh, uh, ongoing efforts right now, uh, among the, uh, industry to have a Gaussian splatting file format standard to, uh, make the, uh, all these Gaussian splatting more available to, uh, you know, all these, uh, you know, clients, many companies, uh, to be able to be, uh, e-utilize Gaussian splats on the web, uh, with a, uh, only a few, uh, you know, megabytes. So that's, uh, something in the, uh, direction, uh, that's been discussed, uh, right now.

Moderator: So is the, is the goal to, to get, uh, to get it working in any device across all the platforms or?

Yoshiharu (Josh) Sato: Yeah. So the... One of the a-advantages of Gaussian splats is that it can be, uh, uh, executed on almost any devices like smartphones, uh, tablets, uh, you know, of course desktops and laptops. Uh, and that's because Gaussian splats is based on conventional rasterization, uh, on the GPU. So it's, uh, rasterization, ra-rasterization, uh, along with alpha blending.

Yoshiharu (Josh) Sato: So there is no, uh, costly neural network or AI. It's just traditional, uh, rasterization in the, uh, original, uh, GPU pipeline. So again, uh, you, you can use it for, uh, phones, uh, laptops, uh, almost any device you have.

Moderator: Thank you. Uh, Tomas, everybody knows you're, you're a photogrammetry king. Uh, if, if 2D is against my religion, I think photogrammetry for you is like a holy grail. Uh, what broke inside of you that you just started using also Gaussian splattings?

Tomas Barnas: Yeah. So, uh, the possibility to combine these two things, because I'm never saying or like, "This is the only thing you're gonna use." We are using also laser scanning and everything, but I still, uh, love photogrammetry as it is for the results you can get. But, uh, generally when you see, uh, what I was showing also in my presentation, that how much you can achieve in a way simpler way. And this is like not about like just trying to do something faster, but it's about the, the possibilities.

Tomas Barnas: So really combining photogrammetry and, and Gaussian splatting into one thing, it's, uh, it's amazing. And, um, on top I would say like... 'cause with, with Michal, we had a lot of discussions always when the NeRFs were there, like the first ones. And I really, like, I was super against because we, we try to use them anyhow. Then putting them in Unreal Engine, it was killing all the GPUs.

Tomas Barnas: It was like super, super heavy for, for anything, and the results were worse. So, so I was kind of like against. But then, uh, yeah, then, uh, the development of Gaussian splats really, really brought it back and, and gave us the possibility to create way cooler things than, than I could ever imagine before.

Moderator: Uh, when the, when the first, uh, first, uh, N-NGP, Instant NGP came and, and the first papers about the NeRFs came, I did a post on LinkedIn saying, uh, in the headline, "Photogrammetry is dead." And Tomas was instantly like, "Michal, don't do that. Please, please don't ruin my business." So that, that, that's why I asked him what, what broke inside him. Uh, guys, I, I'm really glad you're, you're asking the questions.

Moderator: Uh, I would like to touch as many as, uh, as possible. So let's start with the, with the first one. What is the best way to get started with Gaussian splatting? I think, uh, Michal would be, uh-

Yoshiharu (Josh) Sato: Yeah, definitely

Moderator: the, the good choice.

Yoshiharu (Josh) Sato: So in, uh, in terms of, you know, taking a look at getting started with on like the capturing side, you can say... You know, a lot of these datasets will work with existing photogrammetry datasets. So if you already have, uh, something like this, you're able to put it into some of these pipelines, which I'll get to in just a second. But with... In regards to the actual data in which, uh, these different radian field representations really love, is they really love having sharp images and sharp frames.

Yoshiharu (Josh) Sato: So any motion blur that is introduced into a still image, you should actually remove prior to processing it. It actually will do, uh, more to damage the scene than it will to aid. So it's really a, a much better idea to just curate a very well-selected set of images as opposed to just doing more images altogether. Uh, in terms of getting started on the pipeline side, depending if you're looking to train locally, you know, there's some great options like, uh, Postshot, which is free to use in beta right now on Windows, as well as Nerfstudio, which is where I think you can currently get the, the, the highest, uh, fidelity. It gives you a lot more manual control.

Yoshiharu (Josh) Sato: If you're trying to do very large scale captures, then obviously, you know, pos-processing through XGRIDS is an amazing, uh, platform as well. If you're looking to go strictly cloud-based, there are platforms like, uh, Teleport from Varjo, who just introduced their 2.0 version yesterday and you're getting really great results from there. If you have an iPhone or Android device and you wanna just capture everything locally on your phone, you can use, uh, Niantic's Scaniverse as well. And for just getting started as well, I would also suggest you don't need to go and try and capture an entire city or a location all, uh, at the first time.

Yoshiharu (Josh) Sato: You know, maybe you could start small, maybe take a small object, put it on a, you know, little thing like here, and just passively walk around it, take some images, and, uh, put it into one of those, those pipelines.

Moderator: Perfect. Uh, from the very beginning, I think that the most concerns about Gaussian splattings was about the accuracy, about the scale. Uh, but what you did mainly with XGRIDS, uh, did you just decided to, to like, uh, break all the biases around the Gaussian splattings? Because like the combination of, of SLAM and Gaussian splatting, I think it's brilliant. So, so, uh, was, was actually the, the market biases the, the reason why you did this or did you just came up with this idea to, to, to connect SLAM with Gaussian splattings?

Mindy Li: Very good question. So firstly, uh, we came from that surveying industry is our very starting point. And the first thing we must to tackle is accuracy. And then we find the Gaussian splatting itself is really a good way to do the visualization parts. And, uh, but we-- when we combine the LiDAR data together with the Gaussian splats, we also tackle with the things with the accuracies.

Mindy Li: And the first requirements posed by our customer is that, uh, whether we can do measurements based on the Gau- that Gaussian splats. So that's why we combine the point clouds accuracy together with the Gaussian splats. We align them together, uh, accordingly. So that's why we can conduct the measurements based on that centimeter level accuracy, uh, to see that, uh, it's like, uh, if you can, uh, measure a table's, uh, length or height, uh, in a Gaussian splats in a accurate way and easy way, which is totally different from that point clouds do, uh, in some very professional measurement software. So that's the things we try to do, use a more visualized way to do the measurements and try to tackle the b-biases that, uh, Gaussian splats cannot do measurements.

Mindy Li: That's just because that its backs- its backgrounds or the, um, the techniques behind Gaussian splats is in accurate point clouds.

Moderator: Mm-hmm. Mm-hmm. Uh, Michael just, uh, just said, uh, several ways of how you can start actually with the Gaussian splats and, uh, or splattings and, and you can see that you can just pull your phone and, and start with it. So, uh, Josh, I think you must have seen all the possible data sources, uh, for Gaussian splattings. I understand that your, uh, your, mm, task is not to process them, but, but to actually show them.

Moderator: Uh, but since, since you have seen all those possible, uh, uh, sources of Gaussian splattings, what, what do, what do you think, uh, it's actually the question from, from the crowd, what do you think is working, uh, better? Is it images? Is it, uh, videos? Is it 360 videos? What do you think is working best with Gaussian splatting?

Yoshiharu (Josh) Sato: Yeah. So currently, our platform is fully web-based, uh, and it's only for editing and viewing. But our plan is to make our platform fully end-to-end, meaning that, uh, you can not just, uh, you know, edit the scene or viewing the scene, but also you will be able to train Gaussian splat on the web. And in fact, there is already a open source solution. Uh, it's called Brush.

Yoshiharu (Josh) Sato: Uh, it's written in Rust and, uh, utilizes, uh, WebGPU. So it's fully cross-platform, uh, where you can train Gaussians not just on Windows, uh, Ubuntu, Linux, also, uh, Mac. So you can utilize all these NVIDIA, uh, NVI- AMD GPUs as well. So yeah, I think, uh, yeah, I think we are going in the right direction where we can use, uh, you know, all these, uh, you know, different, uh, GPUs or th- or, or these different, uh, uh, you know, uh, machines. So yeah, uh, web, WebGPU-based, uh, platform, it's something, uh, that we are aiming to, um, develop, uh, in the future.

Yoshiharu (Josh) Sato: And as for the, uh, uh, capturing devices, uh, you can use all sorts of, um, co- consumer devices such as, uh, you know, uh, smartphones obviously, and then you can use, uh, 360 cameras. You can also, uh, professional cameras like, uh, DLSR or mirrorless cameras, and then of course professional, uh, LiDAR scanners of, uh, XGRIDS. So there's all sorts of, uh, capturing devices available for, uh, generating 3D Gaussian splats already. And there is no need for, uh, like, uh, you know, special hardware for starting Gaussian splats. You can just start with your own phone, just take a short video around the object, then train Gaussian splats.

Moderator: Michael is eagerly looking at me, just wanting to say something. Feel free guys to jump into each other. Like the, the, the fight is actually the best way for engagement on social media. So let's-

Michael Rubloff: Yeah. So comparing like photo to video for the import, uh, uh, the, the input sources, it really just comes down to where you have the, the sharpest frames. You know, 'cause if you take a video, you just want to really review, uh, and you want to like put into like FFmpeg or get the original like actual input images. You just wanna really cull through and make sure that the images that you are gonna be putting in from that video are gonna be sharp. 'Cause it's very easy to use video as an import, uh, as an input, but just make sure that the actual base images that are being pulled out of that video are still gonna be very sharp because that, that more than anything will, will impact your, your reconstruction fidelity.

Moderator: Thank you for adding that. I, I need to jump there as well. So

Michael Rubloff: You, you go ahead, Eugene.

Moderator: I wanted to exactly say the s- say the same because like that, that's general question and it was many times also around photogrammetry, like images, videos. Why people want to use videos also for Gaussian splats? Because it's easy, straightforward. You don't have to think, you just run. And, uh, the problem generally like Gaussian splats, um, and photogrammetry is about keeping the overlaps and keeping the angles.

Moderator: So what you don't do with video, like with camera, if you're taking pictures, you take a picture of table, then you make five steps and you make a second picture. You ruin the, the, the way how it should be processed. If you make video, you never make that mistake. So the video is helping you to, uh, to keep the workflow it should be, as it should be done, but it doesn't give you better, uh, better results. It just give you like better, better pipelines.

Moderator: So, so the images as itself, they're always, always sharper.

Michael Rubloff: Mm-hmm.

Moderator: Right?

Michael Rubloff: You, you-

Moderator: Yeah.

Michael Rubloff: You want to add something? Uh, yeah. No, exactly. The, the sharper the better. And there's actually a, uh, a platform called Reflect that offers something called sharp frames, where basically you just upload a zip file of it and it'll rank all of your, uh, image sharpness by e- each image, and you can kind of do like auto-culling.

Michael Rubloff: And you could also see in, uh, across like a timeline view of like where the actual images are being removed from, so that if you are missing some level of overlap, then you can kind of just make a choice at that point of what you want to do.

Moderator: Perfect. Uh, Tomas, uh, when I introduced you, I said that, uh, you're one of the first that actually were able to commercially use Gaussian splattings, and there was a question about using it in VR. Can you, can you tell people about the project?

Tomas Barnas: Yeah, exactly. So, uh, as, as Josh was al- was already mentioning, like, you can use Gaussian splatting on the phones on... And, and VR. If you have VR, it's generally like Android applications, so it's, it's, it's the same like a phone more or less. Uh, but before this was even possible, uh, there were still possibility of streaming the Gaussian splatting into your goggles.

Tomas Barnas: So even if the performance or if the headset is not supporting it, uh, natively, so you can't upload it there, you can still make it run on the PC and stream it into VR, which is not a, not a problem at all. So, um, that was also one of the things, like, coming to the first question, what, what made me think of like, "Okay, we need to use it almost in every project because you just combine it with anything you want or you can use it just as it is, like Gaussian splatting." So yes, it's, it's possible to, to use it in VR, and it's a pretty straightforward pipeline.

Moderator: Mm-hmm. Mm.

Mindy Li: Yes. I, I also want to add something. Uh, these days I, I also bring over VR demo here, and, uh, we've already tried to put the same Gaussian splats in Apple Vision Pro, uh, Quest 3, and Pico. So we found that Apple Vision Pro really have a good GPU on board, so that's why his visualization effect is the best. And, uh, we still can use some other streaming methods to do the things, uh, with that Me- uh, Meta Quest 3 and the Pico things.

Mindy Li: So you can also see, uh, it's a very immersive case, uh, with that Gaussian splats 'cause everything is so real. So the, the only thing you need to try-- uh, you need to tackle is that to, to make all the datas more compressed compared with, uh, for the geometry things or the, the conventional Gaussian splats format. Yeah. So it's possible, and it still, uh, have space, uh, to improve.

Moderator: Perfect. Perfect. Uh, Josh, where do you see the biggest limitations of Gaussian splattings ri- right now?

Yoshiharu (Josh) Sato: Yeah, great question. Um, so in terms of quality, uh, of Gaussian splat reconstruction, I would say, uh, yeah, all these, uh, you know, caption methods are, uh, you know, there's motion blurs, um, differences and then the exposures. Yeah. These are all adversarial factors for making high-quality Gaussian splats. So, uh, a technique in photogrammetry, uh, can definitely be utilized in also Ga- Gaussian splatting ca- capturing.

Yoshiharu (Josh) Sato: So yeah, if you have, uh, uh, knowledge and experience in photogrammetry, uh, it's definitely helps in, uh, constructing and generating high-quality, uh, Gaussian splats. So yeah, that's the, uh, uh, yeah, one of the, uh, limitations, uh, for the Gaussian splats when it up- when it up- when, when it comes to generating high-quality, uh, Gaussians. Yeah.

Tomas Barnas: Yeah.

Moderator: Mm-hmm.

Tomas Barnas: And I'd say that one of the historical limitations of, of Gaussian splatting, uh, was basically you, you-- The n- original version of Gaussian splatting, uh, would assume a perfect pinhole sensor. So it would mean that it would really struggle with, like, fisheye lenses and it, it would just kind of like be turned off to that. And so, uh, N- NVIDIA actually released a paper called "3D Gaussian Ray Tracing," which is not Gaussian splatting. It's a pure ray tracing based, uh, radiance field, uh, which, uh, you know, takes away that restriction. Uh, they then came out with another paper called "3D Gaussian Unscented Transform," which actually brings, uh, the secondary lighting effects and pinhole.

Tomas Barnas: Uh, sorry, and fisheye support to traditional Gaussian splatting. So now this has been solved, uh, from these two different papers, and the second one goes into the normal Gaussian splatting. Uh, so that now really opens up a lot more of the use cases for, like, robotics applications, for instance.

Moderator: Mm-hmm. Mm-hmm. Mm-hmm. Um, I think you, you... People that are familiar more with Gaussian splatting, you must remember the first paper from, from NVIDIA.

Moderator: It was called Instant NGP. Uh, and, uh, there was this video with a woman and NVIDIA claiming that you can get perfect reconstruction of the person from only four images, which I honestly think, uh, it was a bullshit. But, uh, it's, it's, uh ... Sorry, NVIDIA. Are you here?

Moderator: Uh, but, uh, uh, my question is, uh, to Tomas because, uh, you've been testing a lot of datasets. Uh, it's coming from here, the question. Uh, do you think it's better to use more images with less quality or less images with higher quality?

Tomas Barnas: Yeah. This is this balancing in between. Like, if you do too less images, then you don't get a good quality. So it needs to be a balance. Like, the amount is, it's hard to say.

Tomas Barnas: Like, you need to do a good amount with good overlap and good quality. Yeah, but like it, it's... No. If we, if we do it in extremes and we say like, uh, overdoing pictures, so making like too much, it's, it's bad for the results, like anywhere. Because if you have like too much of input, it's too mu- too many overlays and, uh, that multiplies possible mistakes or misalignments which are making the, the results worse.

Tomas Barnas: So it really needs to be, uh, like, like a good amount.

Moderator: Mm. Uh, I di- I didn't want to make it sound bad, uh, saying that about NVIDIA because NVIDIA is doing one of the best jobs, uh, with, with all this new, uh, technology. Uh, it was only to, to really point out the importance of, uh, taking the images because, uh, it would be very bad if, if you thought like you can get poor images, poor video or anything, and you will get perfect results. It's, it's always shit in, shit out, right?

Tomas Barnas: Yeah.

Moderator: So what, what's, what's like the, the best workflow do you think, Michael, to just capture some object?

Tomas Barnas: Yeah. I mean, for me now, you know, I, I go back and forth between photo and, and videos, but now I, I just mainly shoot, uh, uh, still images and then, you know. I've seen people

Michael Rubloff: Some extremely high fidelity captures going, uh, just from, you know, normal 2D, uh, mirrorless camera into reality capture and then into either, like, Postshot or, or Nerfstudio, uh, uh, uh, for just, like, general, like, capture, I would say. Um, and then it just really depends on, like, what, what it is you're, you're shooting with. So, you know, for instance, if you're, if you're shooting with your, your phone, then it could be a different pipeline versus with, like, an XGRIDS scanner or something-

Moderator: Mm-hmm

Michael Rubloff: along those lines.

Moderator: Um, Mindy, you must have seen tons of different use cases for, for either your SLAM or also for Gaussian splatting. So where do you see the, the biggest potential? Whi-which industry do you think is going to benefit the most?

Mindy Li: Yeah, that's the thing I, I observe all, observe all the time. So firstly, we find most of the requirements comes from the AEC industries. So that's, uh, that's just because we, uh, use that for spatial modeling. So that AEC industry, uh, so-called, uh, architecture, engineering, and, uh, uh, manufacture... Uh.

Michael Rubloff: Construction.

Mindy Li: Uh, yeah. Uh, so, so that's the things people will buy for 3D asset things, and they want to know exact 3D structures of the whole thing. So, so people also do the scanning and reality captures again and again very often. So, so that's the-- I think it's the, um, very basic, uh, industry who, who needs that Gaussian splatting for, for their reality capture things. Then we, we find the filming industry and gaming industry, uh, increases very quickly.

Mindy Li: And, uh, last month, uh, this month we go to NAB. Uh, it's National, uh, Association of Broadcasting. It's, uh, the biggest, uh, show for the filming and broadcasting industries. We find that we are the first guy to attend this kind of show, uh, to, to that, to do that reality capture things with Gaussian splatting. And from that pre-scouting to the pre-production and virtual production and VFX, all the workflows can use the Gaussian splats to enhance their works- workflows to cr-- uh, increase efficiency, get a better quality, and cut cost.

Moderator: Mm-hmm. Mm-hmm. Thank you. Uh, Josh, uh, you mentioned already that, uh, your goal is to, to create the whole pipeline of, of, uh, working with the Gaussian splattings. Is there on the roadmap also, uh, working with, uh, different, uh, 3D assets just to combining with the Gaussian splatting like, I don't know, like 3D CAD or BIM models or...?

Yoshiharu (Josh) Sato: Yeah. So one of the advantages of g- using Gaussian splats is that since it utilizes traditional rasterization on the GPU, you can use what's called depth buffering for blending or fusing with traditional textured meshes. So you can import, uh, texture, textured meshes like, uh, you know, 4D BIM model, uh, from Revit, for instance. And then you can fuse those, uh, models with Gaussian splats. So it's already possible.

Yoshiharu (Josh) Sato: You can do that, uh, technically. And, uh, yes. And, uh, yeah, it's still, it's i-in our pipeline, uh, in the future to implement s-such functionality.

Moderator: Mm-hmm. Mm-hmm. Uh, Tomas, s-since, okay, I'm saying this third time, but you, you're that, uh, photogrammetry guy. So, uh, please tell me your point of view. How do you see Gaussian splatting replacing or complementing, or h-how do you see the position of Gaussian splatting in, in next years compared to photogrammetry?

Tomas Barnas: I see these technologies merging, simply said. Like, the, the, the target would be then in a certain point you would not really think of that because the pipeline is more or less the same. Uh, the processing is more or less the same. The output is different. So it's kind of...

Tomas Barnas: For me, it's, it's, it's becoming one technology which has just two different results. And I think also the softwares processing the, uh, uh, the photogrammetry will, will be slowly adding all these Gaussian splatting supports, and they will enhance the results. So I think the-- it will end up as, as one, one big thing which will grow together.

Moderator: Mm-hmm. Mm-hmm. Mike, when do you think that the actual hype around 3D Gaussian splattings-

Michael Rubloff: Yeah

Moderator: happened?

Michael Rubloff: When it, when it began? Yeah. So we're actually coming up on the two-year anniversary of the paper's publishing. It, uh, came out in late April of 2023. And for the first couple of months, nobody really paid attention at all to it.

Michael Rubloff: Uh, what really changed was it won Best Paper at SIGGRAPH in 2023, and it was pretty much like overnight the amount of, like, excitement of it just really kinda went through the, the roof. And so the, the two, you know, the, the, the researchers from Inria, uh, George Kopanas and Bernhard Kerbl, uh, they, they released the, the, the code, and it just kinda went from there. So, uh, yeah, August 2023 is when it started to really become popular.

Moderator: Thank you. Mindy, what, what is on your roadmap for, for next years? Where do you see the, the XGRIDS working with the Gaussian splattings? What are next steps? What people can expect from your development?

Mindy Li: Um, so firstly we, we want to make better, um, hardware for easy scannings 'cause that, uh... Always people said that LiDAR is too expensive, but, uh, there are many and many people want to use that for the spatial modeling things, not only to do the objects or the single points. So better hardware. And another things is that the ecosystem things, 'cause we find a limitation of Gaussian splatting, uh, at this moment is that it cannot, uh, well integrated with the traditional 3D workflows very well. But, so these days I work closely with the partnershi- partners like Yoshi, uh, very well and, uh, and Tomas together to see how we can, uh, develop more plug-ins, SDKs, uh, more softwares, and, uh, even, uh, u- uh, uh, uh, implemented, uh, uh, WebGPUs for more workflows integrations.

Moderator: Mm-hmm.

Mindy Li: So that's the, the future chance.

Moderator: Yoshi, if, if you had to pick, uh, just one solution to recommend how to produce nice enough Gaussian splattings, please don't say XGRIDS because it would look bad. It would look like a commercial set, but we don't want, want it to be. So please tell me one, one solution. May-maybe a free solution, like that people can use freely, so.

Yoshiharu (Josh) Sato: Yeah. That's a great question because there are lots of, uh, software solutions or open source projects, uh, that's out there, uh, that lets you generate Gaussian splats already. Um, I would say, um, so if you are just hobbyist or en-entry level user, I would say Scaniverse. Uh, it's a phone app, uh, which lets you generate 3D Gaussians on device without any, uh, you know, requirements for any, uh, cloud, cloud processing or any powerful GPU on the local machine. So you can do, you can generate Gaussian splats on your phone locally, uh, from just a short, uh, video.

Yoshiharu (Josh) Sato: You can just, uh, you know, uh, ca-capture a short video using your own phone, and you can do that, uh, already. So that's a great, uh, solution or app, uh, as an entry level Gaussian splatting, uh, trainer and viewer.

Moderator: Perfect. Tomas, one last question. Uh, uh, you do amazing big scenes, photogrammetry scenes. What is the biggest scene that you actually combined with the, with the Gaussian splatting and, and you saw the perfect results?

Tomas Barnas: Uh, yeah, that's like what I was also showing before. So we had like, um, almost 100,000 images in the scene, which was previously processed without Gaussian splatting. But as it was developed, we just enhanced the, the, the surrounding. And then, uh, putting it together in Unreal Engine created like an endless feeling. So, so that's like, for me, this, uh, this main benefit and, and the scale.

Tomas Barnas: So you, you create like endless scale, and it still looks great because you have it like from the center of the interest.

Moderator: Perfect. Thank you very much, guys. Uh, I hope you get some, got some new information or maybe inspiration. Thank you, Mike Rubloff.

Michael Rubloff: Thank you.

Moderator: Thank you, Mindy Li.

Mindy Li: Thank you.

Moderator: Josh Sato, Tomas Barnas. Thank you.

Speakers

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…

Mindy Li

Mindy LiLinkedIn

Product Manager at XGRIDS

Being the Product Manager of Lixel CyberColor (LCC), a commercial-grade SLAM-based 3D Gaussian Splatting solution, Mindy Li is also in charge of XGRIDS global partnership with strong experience in spatial data capture and immersive 3D workflows, she helps bridge advanced scanning technology with…

Yoshiharu (Josh) Sato

Yoshiharu (Josh) SatoLinkedIn

Founder & CTO/CEO at Waldek Technologies

Is a quant and machine learning expert with a background in high-frequency trading, blockchain, and 3D graphics. He holds a Master’s in Quantitative Finance from the University of Warsaw and studied ML at NYU. Josh has advised hedge funds and startups, focusing on deep reinforcement learning and…

Tomas Barnas

Tomas BarnasLinkedIn

Founder and CEO at Overhead4D and Co-Founder of 3DISE

An expert in reality capturing technologies and the Co-Founder of Overhead4D, a company that specializes in “TURNING THE REAL WORLD INTO DIGITAL EXPERIENCE” focused on blending innovative technologies with creativity. Im an expert member at CIPA Heritage Documentation and passionate about educating…

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

Buy tickets