Gaussian splatting is not enough: 95% still use point clouds – Martin Piala, Blurry

0:00Michal Gula: Martin Piala from, from Blurry is, is our guest. Welcome.

0:03Martin Piala: Hello. Yeah, thanks for having me, guys.

0:05Michal Gula: can you just, like, briefly in, in one word tell us what you do?

0:09Martin Piala: I, I probably need, like, one sentence at least. But yeah. so at Blurry we are building, basically a workspace for reality capture data. think of it as a Google Drive or Dropbox for all your point clouds, Gaussian splats, images, videos all in one place in one bundle where you can analyze them and work with them. That's what we are building at Blurry, essentially.

0:26Michal Gula: wasn't, like, the first, initial idea. Yeah. I, I didn't wanna use the, that wording, but okay. initial idea wasn't about, 3D Gaussian splats.

0:38Martin Piala: It was. It was. I mean, we are still doing 3D Gaussian splatting. we have one of the best browser viewers out there, honestly, for Gaussian splats. But we realized Gaussian splatting is not enough. We need way more than just Gaussian splatting. We need to cover all the current formats and files that users are working with, and point clouds are still not dead. Images and videos are still not dead. They are still being used. so we are trying to figure out how do we put all of it together and use Gaussian splatting as essentially one of the formats, and honestly, one of the ways to nicely present all your reality capture data. But we still need point clouds for measurements. We still need images to have the raw, raw data to, to see.

1:16Michal Gula: I, I remember us having some conversation, like, even few years ago. Like, like, from the very first moment the-

1:22Martin Piala: Yeah

1:22Michal Gula: 3D Gaussian splatting was out there, I, I saw some videos from you and I was like, "Hey, this guy is from Slovakia." "I have to talk to him." So I, I remember we're having a call, and at the first time it was also about the processing of the data, right?

1:35Martin Piala: Yes.

1:36Michal Gula: Towards, 3DGS.

1:37Martin Piala: Correct. Correct. We are, we are doing that now, even, even, even-

1:40Michal Gula: Still?

1:40Martin Piala: Yeah.

1:40Michal Gula: Okay.

1:40Martin Piala: We're still doing that. I think actually the first call we had was actually about neural radiance fields. It was, before Gaussian splatting was even a thing.

1:48Michal Gula: Yeah, can be.

1:48Martin Piala: It was before the... Yeah, yeah. It was, it was, it was-

1:50Michal Gula: NeRFs

1:50Martin Piala: ages ago. Yeah, it was NeRFs. Exactly. So we were doing processing, then we stopped doing that for a while. Then Gaussian splatting came out, and then we realized, oh, now is a good time to actually go into the market with this kind of technology. So, so we revised everything and, and rebuild everything, essentially. So...

2:05Michal Gula: So to make it clear for, viewers and listeners-

2:08Martin Piala: Yeah

2:08Michal Gula: you are not focused only on viewing and displaying and storing, but also-

2:13Martin Piala: Correct

2:13Michal Gula: on processing.

2:14Martin Piala: Correct. Correct. So we do processing now. I appreciate it might be quite difficult to keep track of what we are doing because we are a very young startup, but we move very, very quickly. And a few months ago when we, when we were at 3DISE, it was truly only Gaussian splatting viewing. Now we have all the other kinds of stuff that we did, we do. We have point clouds, we have processing, we have all other, offering that we added since then.

2:37Michal Gula: What, what was the moment for you when you, as a team, when you said or thought to yourself that you need to, like, spread, the, the features inside the viewer? Because I totally understand the, the enthusiasm towards-

2:51Martin Piala: Yeah

2:51Michal Gula: Gaussian splats because it's, it's really looking nice. It's very new. It's hype all over. but was it a competition or was it, like, a deep thinking? Like, what made you-

3:04Martin Piala: Mm

3:04Michal Gula: go and decide that, okay, it, it cannot be only-

3:09Martin Piala: Yeah, so-

3:09Michal Gula: 3DGS.

3:10Martin Piala: I mean, we, we went to the ultimate source, which is our customers and our clients. Like, those are the ones we should be listening to for, for feedback and thoughts.

3:18Michal Gula: I, I hate that.

3:19Martin Piala: Yeah. That's, I... Yeah. But that's, that's how it is. Like, if you want, if you want to build a, a truly useful product you have to listen to your users, to your customers, and listen to their pain, pain points. So we don't really look at a competition, and as much as we can deeply tho- think about our product, that's not enough. So yeah, we just observe the workflows of our users, and we notice that 95% of them are still working with point clouds, and they are kind of like striking into that Gaussian splatting territory and try- trying it out. So we re- realized if we want to capture any meaningful market at all, we have to support what our users use, and then we can possibly help them with Gaussian splats to display everything more nicely. So that was the main reason, honestly. Yeah.

3:59Michal Gula: What I can generally see on the market is that rarely the processing company or, or the company doing processing of the data is... can be as good as, as also in, in a, in a viewing and, and actually using the data.

4:16Martin Piala: Correct.

4:17Tomáš Barnáš: Why do you think you, you can change this n- narrative or, or, or this, this idea that you can be good in both, like the processing and the viewing?

4:27Martin Piala: I don't, I don't think actually we can. That's why we are partnering up with other companies, and essentially our current processing is via Varjo, Teleport by Varjo. You know them?

4:36Tomáš Barnáš: Yeah.

4:36Martin Piala: So... And they are really good at processing, so we partnered up with them, and now we are using their service to do the processing on our end.

4:42Tomáš Barnáš: Oh, okay. Okay, so-

4:42Martin Piala: So we are still focusing on our strengths

4:44Tomáš Barnáš: so, so you are not developing the, the processing part, but you are, like, allowing your users to, to process.

4:49Martin Piala: We actually are, but that's not really public product at the moment. But We will have our own processing soon. but-

4:55Tomáš Barnáš: No worries. We will not tell anybody about it.

4:57Martin Piala: No, no, yeah, I mean, it's fine. It's fine. So we are working on our own processing as well, but as you... Literally as you said, we are trying to solve, like, a very specific problem, not general processing. For that we'll most likely use Varjo for a very long time, but there are, like, certain niche requirements like, oh, we want to process 360 cameras into Gaussian splats, which Varjo doesn't support at the moment, or we want to provide also localized cameras for processing, so we can skip the whole camera localization step in the Gaussian splatting processing. So once again, something that Varjo doesn't offer.

5:26Tomáš Barnáš: But, but Varjo is, is famous for other things, not for the processing of Gaussian splats.

5:30Martin Piala: Oh, yeah.

5:30Tomáš Barnáš: They, they have these cool professional, goggles for, for-

5:33Martin Piala: Yeah

5:33Tomáš Barnáš: VR and AR. So-

5:35Martin Piala: Yeah

5:35Tomáš Barnáš: is there any connection to that, or you just connected with them just because of the processing?

5:41Martin Piala: Just because of the processing, and that's what we do even on, like, a hardware manufacturer side. We are trying to partner up with hardware manufacturers. We have no intention to produce any cameras, LiDAR scanners, that kind of stuff. Like, there are b- other companies that are better positioned to do that, but we would like to partner up with them so that we can ingest their data straight away and organize it and analyze it. Basically do what we are really good at.

6:02Tomáš Barnáš: And do you think it, it would be a problem to try to create the, the connection to their goggles?

6:09Martin Piala: I think there's a whole different can of worms that we have to open if we want to support, like, a VR headsets, just due to performance and the, the fidelity requirements that you need for it to look natural inside goggles, so you don't feel dizzy, you don't feel kind of awkward, you don't, you don't see things that are too blurry. sorry for the pun.

6:29Tomáš Barnáš: Bread-

6:29Martin Piala: You know what I mean

6:30Tomáš Barnáš: bread dropping.

6:30Martin Piala: Yeah, yeah, yeah. So, so that is a can of worms we haven't opened yet just because we re- realize we are not big enough to solve all these problems at once. So first we are trying to solve the problem, like how do we view this comfortably in browser, and then we can possibly look at other devices. So-

6:46Tomáš Barnáš: To, to, to be honest, like, there is no one who solved it yet. And if you are-

6:49Martin Piala: Mm

6:49Tomáš Barnáš: partnering with a, with a Google headset provider, I would just say that maybe that's something what you can just-

6:56Martin Piala: Yeah

6:56Tomáš Barnáš: l- look up because I... It's not a huge market, but once someone will solve it, then there would be products. But-

7:03Martin Piala: Definitely.

7:04Tomáš Barnáš: Of course.

7:04Martin Piala: Definitely. But it's no, it's no easy problem to solve.

7:06Tomáš Barnáš: It's not easy problem.

7:07Martin Piala: I, I looked at it before. The, the, the performance requirements that you have are, are quite steep.

7:12Tomáš Barnáš: De- definitely. I would like to also, tackle a little bit with AI topics because- you are also integrating bunch of AI features. So maybe that's something you can tell us about.

7:23Martin Piala: Yeah, yeah. It's a, it's a very, very interesting thing. So we do have a little bit of AI now, like marketing and branding and positioning online, but we don't actually offer any particular AI at the moment publicly to people. And the main reason why we, why we don't have it yet, even though we are working on it, is that we don't want to put out, like, a half-baked product that is like AI.

7:44Tomáš Barnáš: Mm.

7:44Martin Piala: And you can, you know, ask it where my windows are, but, like, the model is so tiny that you just see where they are. Like, there's no point in having the AI for, like, you know, like a toy, a toy experiment, like a toy prototype, and that's why we are building this whole, Google Drive or Dropbox for reality capture data. We take all the data, we work with it, and that's how we want to build the context for the AI to do something meaningful. So we are building it from the, from the ground up in a meaningful way, meaningful fashion. Once we have all the data, then we can do something smart about it.

8:14Tomáš Barnáš: The difficult situation is that, different types of the data, are developed or, or created in different, times and different, like, timelines in, in, in the-

8:24Martin Piala: Yeah

8:24Tomáš Barnáš: like, overall life of the building, for example. So it can e- easily happen that the, I don't know, like, BIM documentation can, have window somewhere-

8:34Martin Piala: Yeah, yeah

8:34Tomáš Barnáš: where the scan it doesn't have.

8:36Martin Piala: Yeah, correct.

8:36Tomáš Barnáš: the CAD drawing can have doors somewhere else or a wall somewhere else. So, does that mean that the AI will also be able to identify these discrepancies in?

8:46Martin Piala: So I think we, with the current frontier models, especially Astra from OpenAI that just came out, or Fable, these models are pretty smart, and, and what you can do as a human, these models can do. And if you as a human naturally understand, okay, so this model is timestamped, you know, three months before this particular BIM scan or BIM model, you can naturally tell that, oh, this is probably more outdated. This is the fresher data. And the same thing, these, like, AI models can do as well. Like, as lo- as long as you expose all of this information, as long as you don't, like, drop them in blindly into the model and, like, tell them figure something out, if you give them the information, "This model is from August 2026, but this one is from September 2026," they will, they will get it. Like, it's... It is not a, it's not a rocket science as a-

9:31Tomáš Barnáš: Since you mentioned the As- Astra, like, the videos I'm, I'm, I'm seeing on the, al- all over the internet, I, I reposted several of them. Like, my brain stopped braining when I saw that. Like, are, are you, like, actively testing and, and using Astra?

9:44Martin Piala: Yeah, yeah. I mean, we use Astra for almost for e- every... I mean, Astra and all other models. Like, we use all of them. We use, you know, Astra, Fable, from ZAI, the GLM models. Like, everything, we try all of them essentially 'cause we want to be on t- top of our game. And since we are a small startup, we will use any help we can get. But I do agree, they are really impressive in what they can achieve, but you still need this, like, this, like, human hand-holding every now and then. So, so yeah. We can't just let them run freely, but they are pretty good, and they are definitely making us way more efficient in, in our work.

10:17Tomáš Barnáš: Do you have to pay for all the models? Or-

10:19Martin Piala: Yeah

10:19Tomáš Barnáš: as a startup, are, are you in some incubator or some, some support do-

10:23Martin Piala: No

10:24Tomáš Barnáš: Are you, are you having some, like, incentives for-

10:26Martin Piala: not necessarily. We, we, we pay for all our, like, access. I mean, we have our subscriptions and, and everything. We do have some favorable pricing-

10:34Tomáš Barnáš: Mm

10:34Martin Piala: every now and then. but yeah, we do pay for, for, uh- 4DX is the same as other, other companies do

10:40Tomáš Barnáš: You, you also mentioned that, like, the AI detection would be interesting, but with the smaller models you can still do it by your own, like you said. So, are you limited to a size of the model right now, or is it, like... Because you mentioned, like, smaller models. So how about the scale? Like your users can detect-

10:59Martin Piala: It's a, it's a good question. I don't think that is the... The size of the model is not the limiting factor. me, you know, putting on the engineering and AI engineer hat right now, the, the limiting factor is the data format at the moment.

11:12Tomáš Barnáš: Mm.

11:12Martin Piala: these models, they don't really understand, like, a 3D data or point cloud. So like the 3D geometry as it is. They are much better at, at understanding, for example, images. So instead of giving the full model, you might want to give it screenshots or recordings of your, of your data- because they will understand it naturally a bit more, which is interestingly how us humans interact with all this data anyway. Like, we don't really interact with it as, like, a 3D space, or unless you put on a VR headset, we still view everything through our 2D screen. So, yeah.

11:44Michal Gula: But, but this, this probably is just the case of, of, like, not enough training data.

11:50Martin Piala: Exactly.

11:50Michal Gula: Because if, if there is one crazy company- that could take, like, millions of square meters- scanned with, like- any scanner, put them into a training model-

12:00Martin Piala: Correct

12:00Michal Gula: like, this will be solved.

12:01Martin Piala: Yes, this will be-

12:02Tomáš Barnáš: You could actually take the entire Earth now, because everything-

12:04Michal Gula: Yeah

12:04Tomáš Barnáš: is more or less scanned, right?

12:05Martin Piala: And I... E- exactly, exactly. You... What, what you said is precise. We just don't have the data to have these models. That's why we have to-

12:11Michal Gula: Oh, we do have the data.

12:12Martin Piala: Yeah.

12:12Michal Gula: We don't have that crazy person that would do that right now.

12:14Martin Piala: Oh, yeah. Yeah. Not-

12:15Michal Gula: Yeah

12:15Martin Piala: yet. Not yet. Actually, you know, I would argue we don't have enough data. I wouldn't under- underestimate, like, how much text data is out there on the internet, and how does it compare with a, with a, with the 3D scans, because it's much... It's cheaper to produce text, and there is loads and loads of it online. That's why LLMs are really, really good, these large language models, because they're trained on loads of English and actual international texts, but, uh-

12:36Michal Gula: Y- you know, but the f- the funny part is that the-

12:39Martin Piala: Mm

12:39Michal Gula: the, these AI models, they don't actually have to understand the 3D data and work- with the 3D data be- because at the end, it's all text. It's TXT.

12:48Martin Piala: Ultimately, ultimately, yes.

12:49Michal Gula: It's just a database of points, colors-

12:52Martin Piala: Yes. Yeah

12:52Michal Gula: right?

12:53Martin Piala: Yeah.

12:53Michal Gula: So these, these are, like, only words, letters, and numbers, which might be actually much more easily to, to work with compared to, like, understanding the 3D views and everything else.

13:05Tomáš Barnáš: I just got-

13:06Martin Piala: Yeah

13:06Tomáš Barnáš: got scared that you will say, like, they-... they don't need to understand 3D data because on the end everything is just 2D, so...

13:12Martin Piala: Yeah. Whoa, whoa.

13:12Michal Gula: No, that, that's against my religion.

13:14Martin Piala: It'll be too much. It'll be too much. Let's not go that far.

13:17Michal Gula: Don't, don't push it.

13:18Martin Piala: Yeah, yeah. No, don't. Yeah, yeah. No, no, you are, you are, you are ultimately right, but there are, there are tricks to make it easier for them to understand. The same way you don't read a huge text file full of, you know, numbers that describe your point cloud. You look at it via a screen. We can, we can make it easier for, for these language models and AI model as well. So...

13:36Michal Gula: And, and maybe just to prove myself wrong- Like, like the next step after the language models, like obviously there were, visual models, right? So... And these are working, like pretty well.

13:49Martin Piala: Yeah, they are working pretty well. Yeah.

13:49Michal Gula: Like you, you have AIs that can, like you already mentioned that- they can easily like identify windows, doors and-

13:55Martin Piala: Yes

13:55Michal Gula: and everything else.

13:55Martin Piala: Correct.

13:55Michal Gula: So, and this will be only getting better. So I might have actually said, a stupid thing with the, with the... with the text, editors and, and databases because, even Tesla is doing this, right? They-

14:09Martin Piala: Exactly

14:09Michal Gula: they are everything identifying and working with the, from the visual side of view.

14:14Martin Piala: Exactly. You w- you, you want to feed them what is easiest for the AI model to recognize and understand.

14:19Michal Gula: Mm.

14:19Martin Piala: And image is actually pretty good because once again, we have loads of images online, so it's like a huge training data set essentially that we can use. But the amount of 3D data is still much smaller than, than all the images online essentially. So...

14:34Michal Gula: You mentioned that you basically, like this pivot- was created like half year ago, or...?

14:42Martin Piala: You mean pivot towards like the whole-

14:43Michal Gula: Oh, the, the, the blurring

14:44Martin Piala: Cubox?

14:44Michal Gula: The blurring.

14:44Martin Piala: I wouldn't necessarily call it pivot. It would be... I like to refer to it as a course correction. Like

14:51Michal Gula: You, you said that the company was created like half year before the 3DISE.

14:54Martin Piala: Yes.

14:54Michal Gula: So, and w- we met like three years ago-

14:58Martin Piala: Oh, right, right

14:58Michal Gula: so, so that's why I'm, I'm trying to say that-

15:00Martin Piala: Oh, I see

15:00Michal Gula: it was a pivot, but okay.

15:01Martin Piala: Oh, okay. I, I see what you mean. Yeah, I mean, back then it was even called differently. It was a different company, different technology. The, the, the research wasn't there essentially. But yeah, but that, that's why we try, we kinda like revived everything and started again-

15:13Michal Gula: Okay. So, so I, I want to get to, to the question, like, you are early startup right now.

15:18Martin Piala: Exactly, yes.

15:18Michal Gula: ha- have you already like get some like angel funding or...?

15:23Martin Piala: Yeah, yeah. We, we got pre-seed funding from two VC funds, so we are pretty, pretty well set up for, for our start.

15:30Michal Gula: When, when do you expect the next round?

15:32Martin Piala: it's a tough question. It's a tricky question. Yeah.

15:34Michal Gula: Be- because once you hop on a train, like you gotta go-

15:37Martin Piala: Once you hop on a train

15:38Michal Gula: all the way

15:38Martin Piala: Yeah. Once you hop on a train, you gotta... Exactly. Exactly. So we would hope for the next rounds to happen within the next six to nine months. I think it'll be good.

15:47Michal Gula: Will it be... It, it, it will probably not be the A round, right? It will be like a s-

15:49Martin Piala: No. The next-

15:50Michal Gula: proper seed round, right?

15:51Martin Piala: Next round will be a proper seed round. So our current runway is around, around three years essentially. That's what we can live with, but we don't want to wait until the last moment to raise because as you said, it's, it is a train. We need to hop on it and we need to, you know, bolt to the destination as quickly as possible. So that's why we are hoping for the next round in six to nine.

16:08Michal Gula: Well, tho- those guys are giving you money for a reason, right?

16:09Martin Piala: Exactly. Exactly.

16:10Michal Gula: So you, you need to play along. Uh-

16:12Tomáš Barnáš: Are you set now in Europe or US?

16:14Martin Piala: So we, our headquarters is, is in London.

16:17Tomáš Barnáš: Okay.

16:17Martin Piala: So literally in the middle.

16:19Tomáš Barnáš: Yeah. Okay.

16:19Martin Piala: Between

16:20Tomáš Barnáš: Yeah.

16:20Michal Gula: So during the seed run you will be aiming for few millions or h-

16:24Martin Piala: Yeah

16:24Michal Gula: higher hundreds?

16:25Martin Piala: Yeah. I don't, I don't... Few millions definitely. It'll be, it'll definitely be in the millions. I don't know how much. That's information we don't possess at the moment.

16:33Michal Gula: Fair.

16:33Tomáš Barnáš: No,

16:33Michal Gula: of course. I, I, I just want to create the image for, for the viewers, for the audience-

16:38Martin Piala: Yeah, yeah. Makes sense

16:38Michal Gula: because like, a lot of companies, a lot of, important people are watching this and, and listening to it.

16:44Martin Piala: Yeah.

16:44Michal Gula: So I just, would like to, to create the, the perception of how you look at it and what's expecting in the next steps so that-

16:52Martin Piala: Yeah, I mean-

16:53Michal Gula: can

16:53Tomáš Barnáš: potentially get in touch with you.

16:54Martin Piala: Yeah, I mean that'll, that'll be cool. I mean, I'll be honest, the market is, is big enough and this conference itself is a proof of that. So the market is huge and it requires a big investment to even, even start to compete with, with, with big incumbents. So yeah, definitely, definitely millions. That will be our seed.

17:11Tomáš Barnáš: Yeah. Yeah.

17:11Martin Piala: Yeah.

17:12Tomáš Barnáš: All right.

17:12Michal Gula: Great.

17:13Tomáš Barnáš: Yeah. We just, we should-

17:13Michal Gula: Thank you very much

17:14Tomáš Barnáš: end live with that.

17:15Martin Piala: Cool.

17:15Tomáš Barnáš: And, thank you for joining us.

17:17Martin Piala: Yeah. Thanks. Thanks for having me, guys. Yeah. Thank you very much.

17:19Tomáš Barnáš: Good luck, and see, see you at, 3DISE next year, right?

17:21Martin Piala: Cool. Cheers.

17:22Tomáš Barnáš: Yeah. See ya.

17:23Martin Piala: See ya.