Lukasz Mirocha of Hong Kong-based Opus Lab shares the lessons of three years of Gaussian splatting through two projects. Liminal Constellations, a 3DGS heritage project built in game engines and LED volumes, worked technically but reached almost no one; its successor Threads of Color, premiered at 3DISE, combines a SLAM capture of Hong Kong streets with an AI-generated underground level and will run in a browser, built in public.
“if you can't just share a link to your experience and you rely on complicated setups, any kind of public pre-presentation gets com-complicated.”
“Don't cling to a tool just because it's a familiar tool. Try to future-proof your tool chain.”
“the cost of learning a new pi-pipeline is lower than the cost of building something nobody can actually reach.”
Lukasz Mirocha described it as an Opus Lab browser experience powered by 3DGS: a SLAM capture of streets in Hong Kong's Soho and Central West above, an AI-generated underground network below, with street murals as portals between them.
According to Lukasz, that is the biggest choice to make; local processing has become much easier thanks to new optimization and compression tools, so you don't have to stick to the cloud.
Lukasz's lesson is not to cling to a familiar tool but to future-proof the toolchain, because learning a new pipeline costs less than building something nobody can reach.
Transcribed automatically (ElevenLabs Scribe) and lightly edited. It may contain small errors. Click a time to play the video from there.
Lukasz Mirocha: Okay. Testing, testing. Okay, you can hear me. Thanks for sticking. It's, uh, getting late, but thanks for staying with us.
Lukasz Mirocha: So three years of splatting with 3D Gau- Gaussians. It's gonna be two kinds of stories co- co- combined into one, so let's start. So I've spent the last 10 years actually hacking, uh, tinkering, pushing the game engines to the limit, so Unity, Unreal. This is my com- comfort zone, and this is very, very im- important in relation to the, uh, rest of the presentation. I'm also a consultant and an educator across Asia and Europe 'cause we, Opus Lab, a company that's pushing the limits of real-time 3D, f- is actually based in Hong Kong.
Lukasz Mirocha: So my 3D Gaussian story. So it's been three years, and some of the projects were my solo projects. Some, some of the other projects were under the Opus Lab umbrella. So we've spent two-plus years really hacking, testing, checking all the, uh, 3DGS pipelines we could find. And as a result, uh, we designed four u- university classes, more than a dozen industry workshop, uh, worldwide.
Lukasz Mirocha: Uh, also as for the... 'cause community is super important for this event, but it's also super important for us. Uh, so together with Mi- Michael Rubloff, we've been organized sessions at, uh, SIGGRAPHs, both SIGGRAPH Asia and the main SIGGRAPH. I've been also an advisor for ARC 3- 3D, and I've also contributed to, uh, Gaussian Splatting, Gaussian Splatting Handbook 2026 by Gabriela. If I was to translate all this, all the previous slides that were mostly text into videos and visuals, this would be this in a nutshell.
Lukasz Mirocha: So we've been testing, we've been hacking, we've been comparing for the last two years. It's all, it all looks nice. You can find it online, so let's not waste time on this, uh, because this presentation is not a tool review. Far from it, a- actually. So it's about the lesson I learned the hard way, and actually, I'm giving it for free to you 'cause it might save you two years, and potentially some money as well.
Lukasz Mirocha: So let's roll. And the clicker is not cooperating. Now it's collaborating again. Okay. So, uh, Liminal Constellations.
Lukasz Mirocha: Please raise your hand. Who has heard about this project? You are cheating, but it's okay. Uh, so this is a 3DGS project that almost nobody has seen, or almost nobody has heard about, and this is exactly kind of a problem indeed, yeah, 'cause you make some- something and then it's really hard to show. So let's start with this one.
Lukasz Mirocha: So what was the aim of the project? So we really wanted to, uh, combine real, uh, real world SLAM/RG- R- RGB capture with procedural real-time cinematography within a game engine, and of course, design some interactive immersive spa- spaces as a byproduct of that. So in essence, what that, what does it mean? So we were interested in testing 3- 3DGS within Unity and within Unreal. Uh, of course, 3DGS is there.
Lukasz Mirocha: Of course, some LED volumes, uh, so LED volumes and some XR devices as well. And one of the aims was to pro- propose a new, uh, refreshed framework for, uh, dig... for culture heritage pre- reservation using those new tools. And pri- primarily, this project has focused on a historical mansion built in 1930s in Hong Kong, which is in a state of semi-decay, so it was also a pre- preservation aspect 'cause this weather is not gra- gracious over there to good old concrete. So project goals again.
Lukasz Mirocha: What was on our checklist? So 3DGS, Unity, Unreal com- compatibility. Tested, checked. We checked all the... Please do take photos.
Lukasz Mirocha: It's all o- o- o- obsolete. Be my guest. It's all obsolete, uh, 'cause it, it moves so fast. Uh, of course, we also wanted to test, uh, 3DGS capturing methods, so from your simple iPhones through DSLRs, SLAM scanners la- later got on board, then some drones as well. So we just threw everything at it.
Lukasz Mirocha: Poor mansion. Uh, so but the mansion has been preserved. At, uh, and as a bonus, the project demo has been presented at a high-caliber event at, uh, SIGGRAPH in Van- Vancouver last year. So practically success, yeah? We can all have a beer and just, just get on with our, with our lives.
Lukasz Mirocha: But there's one problem. Uh, where is the audience for that? And this is exactly a little bit problematic, so let me zoom in on that. So a link, no problem. Just bring your workstation, preferably forty-eight gigabytes of VRAM if you can.
Lukasz Mirocha: You can also throw in a spare LED volume if you have one in your be-bedroom, of course. It's all fine. Then you can see it. Then, then you can watch it and then you can enjoy it. So what are the lessons that I got from doing, uh, Liminal Constellations around twenty twenty-four, twenty twen-far-five?
Lukasz Mirocha: So if you want a high, high craft, familiar tools, and low reach, that's the way to do it. Keep on rocking, yeah? It's gonna be awesome. S- But also on the, on a more serious note, so academic slow timelines and funding and research re-reviews, it doesn't go well. All, all those factors, it...
Lukasz Mirocha: they don't go well with very rapid evolution of tools. And also, this is crucial, I think, bending a new tool and just trying to put it inside a container you, you know, so familiar game engines and all those L-LED vo-volumes, all the cool toys you usually have, might, might be a nice way, a first approachable way of doing things, but it might, might not be the best one, actually. And on top of that, if you can't just share a link to your experience and you rely on complicated setups, any kind of public pre-presentation gets com-complicated. So, for example, the Hong Kong premiere is still pending. Last month pre-premiere in D- in Dubai, well, you know how this ended, yeah?
Lukasz Mirocha: Okay. So it's twenty twenty-three, four, now it's twenty twenty-six. So the big question is, th-that the tools, they moved on. But the question is, did we? So if we look at a thr- at a 3D, not even 3DGS, but at a 3D la-landscape as such, we see a very interesting con-convergence hap-happening.
Lukasz Mirocha: So the real-life cap-captures, they're s- they're starting to con- to converge with synthetically generated 3D da-data. So from one side, we have all those real-world cap-captures that you guys are mostly, are mostly in-interested here. SLAM, scanners, photogrammetry, all this good stuff, and accurate. But on the other hand, you have all those synthetic, uh, AI-generated visuals. So you have world models, you have tools like Marble by, uh, Wor-World Lab.
Lukasz Mirocha: So you have a new ecosystem of tools that actually don't need any kind of real-world capture, just a prompt, either a text or a simple image. So what if we can combine this? And this is actually... So it's not only XGRIDS that had a premiere in the morning, it's also us. That's the world pre-premiere of Threads of Color.
Lukasz Mirocha: So you're the first pe-pe-people who are actually hearing this. And thanks to this, we have a five hundred percent raise in awareness on the content we are m-making in comparison to the previous pro-project. So, great success. So it's gonna be two kinds of spatial memory, but packed as one experience, actually. So it's gonna be a spatial e-exploration of urban, um, urban areas in downtown Hong Kong.
Lukasz Mirocha: Very, very dense city grid. You cannot ima-imagine. And it's all gonna be powered by 3DGS. On the surface, it will be based on a photorealistic SLAM ca-capture of Soho and Central West, two districts in Hong Kong. And on the underground level, it's gonna be about AI synthesized dungeon ne-network built between the streets, and this level will be purely AI gen-generated.
Lukasz Mirocha: On the game mechanics, it's gonna be more about like walking si- walking simulator meets treasure hunt. Um, and street mu-murals will be portals between those two spaces, those two levels. Yes, it will run in a browser. No LED vo-volume needed anymore. No powerful workstation anymore.
Lukasz Mirocha: And yes, you will be able to play it as well. So if we look at Liminal Constellations, very nice solo, more of an academic project here. Very interesting to... in... interesting tech or m- or maybe even art demo.
Lukasz Mirocha: Uh, but when, when we look at Threads of Colors, actually, it's gonna be more of an audience-oriented and ecosystem-enriching pro-project, and it will be built in public. So no more secrets, no more wai-waiting for, for peer rev-reviews. You will be able to witness it, 'cause we will be building it in public. And what connects those two projects is actually that both are e-exploratory, both are about technology or tool, tool chain validation. It's about mapping the horizon of what is possible actually these days.
Lukasz Mirocha: And Threads of Color, two kinds of spatial memory, one experience. Let me introduce the project part-partners here. So of course, it will be developed by us, by Opus Lab. But the, uh, s-strategic funding par-partner will be Vi-Viverse. It's an HTC divi-division.
Lukasz Mirocha: And they will provide all the, all the content ho-hosting, user access, of course, and, uh, some development support as well, 'cause it will have to run in a browser. Yeah, we've seen that logo many times today, but never, never enough. So as for the surface level, for the real-life da-data capture, we will be... we actually have lev-leveraged XGRIDS' por-portfolio of devices, both the, uh, PortalCam and L2 Pro. It will be all processed lo-locally on s- on state-of-the-art Dell, uh, workstations po-powered by, for example, ni- 96 gigabytes of VRAM in an NVIDIA RTX Pro.
Lukasz Mirocha: And I cannot announce it today, un-unfortunately, but there will be another, uh, part-partner for AI syn-synthesis and for some, uh, de-de-development stack support as well. Of course, for initial location scou-scouting, we would leverage traditional 3DGS tools, like mentioned, the Lickfields 2, Studio Play, Canvas of course, and Spark. And it-- if everything goes well, that's the plan, you will be able to play it before Christmas this, this year. This year. This year.
Lukasz Mirocha: That's quite important to stress this. So where are we right now with this? Where should I stand? Here. So the SLAM-based 3D ca-capture and, uh, processing was actually taking place quarter one of, uh, twenty twenty-six.
Lukasz Mirocha: So what was, what was captured? Actually, uh, a dense city grid ca-capture was performed with the, uh, L2. It's a little bit more than, than one square kilo-kilometer. A mid-size object, so the said mu-mu-murals. More than 15 murals were ca-captured with the, uh, PortalCam 'cause we were after the highest fidelity possi-possible that can come from a SLAM sca-scanner.
Lukasz Mirocha: Uh, some da-data processing and, uh, cleanup, so of course SuperSplat and the to- the tools that the 3DGS community knows well. And sur-surface level was blocked for reference, so that's the animation that you just saw. So three years of splatting with 3D Gaussians. Three years and three lessons. Don't cling to a tool just because it's a familiar tool.
Lukasz Mirocha: Try to future-proof your tool chain. Try, try to explore beyond it. And the cost, uh, the cost of learning a new pi-pipeline is lower than the cost of building something nobody can actually reach. Of course, it all depends on the project, on the audience you have in mind, but this is a lesson that I really learned the har- the hard way. And of course, find the room where the right people and partners are, and I think we are all in this very room.
Lukasz Mirocha: 'Cause I met many of you actually last year here at 3DISE. Th-this is my second time. And of course, follow us. Uh, we'll be sharing everything, so the things and the pipelines and the discoveries that lead to success and our fai-failures as well. And the best way to, uh, actually stay in touch with us is, uh, just to follow me on LinkedIn.
Lukasz Mirocha: And that's the slide you should take a photo of. And that's it. Thank you very much for your attention.
Host: Lukasz, thank you so much.
Lukasz Mirocha: Thank you.
Host: Thank you for sharing your lessons. We've got a couple of questions. Uh, let's put them up over here. So, uh, do you see similarity in Liminal Constellations project with Cyberpunk 2077 braindance?
Lukasz Mirocha: So there is a general pro-problem with Cy-Cyberpunk vibes and, uh, H-Hong Kong because it's actually when it rains a li- a, a little bit, there's high hum-humidity, so we actually live in Cyberpunk al-al-already. When you have all those neons and the mist and the chaos on the streets and, and everything. So, uh, it's-- the similarities are deeply embedded in my mind, so I cannot escape it. We live in a cyberpunk world.
Host: Thank you. One more question. What is the biggest pain point for creating 3D, uh, 3D Gaussian splats today for you?
Lukasz Mirocha: I guess it really depends whether you wanna go the cloud processing way or the local way. Uh, this is the biggest quest-- the biggest answer you have to find. And if you go the local way, of course, there's lots of optimizations, there's lots of com-- new, uh, compression tools that are ava-available. So actually, local compression got, got much easier, and you don't have to stick to the cloud proces-processing. That would be my, my main concern.
Host: Thank you so much. Let's give it up for Lukasz one more time.
Lukasz Mirocha: Thank you.
Host: Yes. Thank you.

Founder of Opus Lab
Lukasz is a spatial computing strategist and Founder of Opus Lab, a company focused on advanced 3D, XR, and Generative AI workflows for enterprise and creative industry clients. His work spans strategic consulting, spatial media production, and the development of next-generation 3D pipelines, with…