Table of Contents Show
Fernando Rivas Manzaneque (LinkedIn) co-founded Volinga AI after completing doctoral-level research in machine learning and radiance fields at the Arquimea Research Center and the Universidad Politécnica de Madrid – see his references on Google Scholar). Under his leadership, Volinga has positioned itself as one of the first companies to offer a production-ready solution for 3D Gaussian Splatting (3DGS) tailored for VFX, virtual production, and real-time workflows, most notably with the Volinga Suite and its companion plugin for Unreal Engine.
DP: Hi Fernando! I made, like, 2,000 pictures of my room. What can I do with that?
Fernando Rivas Manzaneque: Good news: those 2,000 images are basically a goldmine. With 3D Gaussian Splatting, the technology at the core of Volinga, you can turn that pile of images into a fully navigable 3D replica of your room, and you don’t need any 3D background to do it. No manual modelling, no UV unwrapping, no texture painting. You just feed the images in, and the algorithm figures out the geometry and appearance from scratch.
What you’d actually do is load those shots into the Volinga Suite, let it process them, and within a reasonable amount of time, you’d have a 3D Gaussian splat of your room ready to go. From there, you pull it into Unreal Engine with our plugin, relight the whole thing, drop in some sci-fi assets, and suddenly your bedroom is a space station. Or a crime scene. Whatever you’re into.
Here’s the honest part, though: your first result will probably be good enough to get you genuinely excited, and not quite good enough to leave you satisfied. That’s the Gaussian Splatting rabbit hole right there. The entry point is surprisingly low. Getting started requires almost no 3D background. But the ceiling is very high, and the gap between “cool demo” and “hero set on a Hollywood production” is mostly about understanding the capture itself: how you move, how you shoot, how you deal with light and motion. Once that clicks, you can build photorealistic environments that would normally require a full art department, and you can do it without leaving your living room.
DP: Okay, so I load my pictures into Volinga, and Unreal shows me my room on a giant LED wall?
Fernando Rivas Manzaneque: Pretty much, yes. You load your images into Volinga Suite, our desktop application for creating and editing Gaussian splats, and the software processes them into a 3D asset you can take anywhere.
The output can come in a couple of formats. The most widely used right now is PLY, which is essentially the format that emerged from the original Gaussian Splatting research. It was convenient, it was already there, and the community adopted it quickly. We’ve also developed our own format, NVOL, which carries additional metadata that becomes genuinely useful once you’re inside Unreal Engine: things that PLY simply wasn’t designed to hold.
From there, yes, you can absolutely pipe that into Unreal and send it out through a giant LED wall. I’m glossing over a few things here, learning Unreal, setting up the hardware, the whole VP infrastructure, but that’s not really the Volinga part of the equation.
If you already know how to work in Unreal, and I suspect most people reading this do, then bringing in a Gaussian splat with our plugin is as simple as drag and drop. With the Volinga plugin, 3DGS assets are first-class citizens in Unreal Engine. They sit in the content browser, they respond to the scene, and you work with them the same way you’d work with any other asset in the engine.
What even is Volinga?
DP: So the Volinga Suite does exactly what with my files?
Fernando Rivas Manzaneque: Under the hood, Volinga Suite runs through several distinct stages. First, it locates your images in 3D space. Then it applies lens distortion correction. From there, it generates a sparse point cloud. And once that foundation is solid, it trains the actual 3D Gaussian Splatting model on top of it.
That first part, the image localisation, the distortion correction and the point cloud generation, is what’s known as Structure from Motion. This is where the algorithm goes hunting for patterns across your images: visual features it can recognise in multiple shots from different angles, which it then uses to triangulate positions and solve the combined puzzle of camera placement and lens calibration. This is also where things can go wrong.
Two of the most common culprits are featureless flat surfaces and insufficient overlap between images. If your scene has large untextured areas, a white wall, or a polished floor, the algorithm simply has nothing to grab onto. And if your shots don’t share enough visual information with their neighbours, the whole chain of correlation breaks down.
Shooting technique matters enormously here. What you’re after is parallax: you want to see the same features from genuinely different viewpoints so the geometry can be triangulated properly. That means moving with the camera in translation, not just rotating on your own axis. A common mistake is to stand still and pan. Keep your feet moving, Shift your position, cover the scene from angles that actually differ in space.
Beyond that, a few things will quietly destroy your results. Blurry pixels, whether from motion blur or shallow focus, are trouble. Lens flare is trouble. The reason is the same in both cases: these are single-image phenomena. They don’t exist in 3D. They can’t be explained across multiple views, so the model will try to accommodate them anyway and produce artefacts in the final splat.
The same logic applies to anything that moves during capture. People walking through the frame, objects shifting, and even abrupt lighting changes. 3D Gaussian Splatting models only three spatial dimensions. Time is not part of the equation. Anything that changes between frames ends up as what we call floaters: ghostly smears of geometry that appear to hang in the air with no explanation, because the model invented them trying to reconcile images that should agree but don’t.
If you want a field-tested breakdown of all of this with patterns, environment types and camera settings, one of our artists put together a practical guide on the Volinga blog: Capturing Gaussian Splats: Lessons from the Field.
An Introduction to Gaussian Spalts
DP: We keep saying “Gaussian splats,” but what does that actually mean?
Fernando Rivas Manzaneque: Gaussian Splats don’t exist in isolation. They belong to a larger family of techniques called Radiance Fields, which has been an active area of research for years. The field really exploded in 2020 with the arrival of NeRF (Neural Radiance Fields) and hasn’t slowed down since.
To understand why any of this matters, it helps to understand what came before. Until recently, almost every reality capture pipeline was built on the same foundational idea: triangulate points in space, connect them, and generate a polygon mesh that represents the geometry of a real scene as accurately as possible. And it works, to a point. The problem is that polygons were invented for modelling the world, not capturing it. They’re a tool designed to give artists control over geometry.
When you flip that purpose around and ask the algorithm to reconstruct reality, you often get results that are decent but rarely ready to use: holes, artefacts, and geometry that needs manual cleanup before it goes anywhere near a production pipeline. That’s not a failure of the tools. It’s just a mismatch between what the tool was designed for and what we’re asking it to do.
Radiance Fields come at the problem from a completely different direction. Instead of trying to build a mesh, they represent a scene as a radiance field: a description of the light emitted from every point in space, and how that light reaches any given camera. The world, in this model, isn’t made of polygons. It’s made of millions of infinitesimal particles that emit light, and the goal is to figure out what those particles look like and where they are.
Here’s an analogy that might make it click. Imagine you’re trying to map the temperature of a lake: not just the surface, but the full three-dimensional volume, deep enough to understand how currents flow. You do it by placing a thermometer at hundreds of different points, at different depths, different locations. The more measurements you take, the more accurately you can reconstruct the temperature field, and the better you understand the currents. You never directly observe the full field. You sample it.
That’s exactly what we’re doing with cameras and Radiance Fields. The camera is the thermometer. We can’t directly measure the light emitted from a single point in space, but we can measure the light that reaches the camera sensor from every point it can see. The more cameras we use, from more positions, the denser our sampling of the radiance field, and the more accurately we can reconstruct it. That’s also why Structure from Motion matters so much: before we can interpret what any camera is seeing, we need to know precisely where in space that camera was sitting.
Now, where does the machine learning come in? The radiance field is an extraordinarily complex mathematical function. Machine learning is, at its core, an extremely powerful function approximator. You give it a set of measurements of some unknown function, and it builds an estimate of that function that fits the data. That’s the NERF approach: a neural network that you can query at any point in space to ask what colour and intensity of light would be emitted there.
Gaussian Splatting changes one fundamental thing in this picture. Instead of a neural network that you have to query for every point in space, including all the empty ones, you get a set of discrete elements, the Gaussians, each of which stores the radiance information for its own region directly. Each gaussian has a shape, a colour, an opacity, a rotation, and these properties are optimised during training until the collection of gaussians together best reproduces what the cameras saw.
Critically, Gaussians exist only where the scene has content. Empty space has no Gaussians. With NeRF, the network had to store information about both the void and the matter. With Gaussian Splatting, you only represent what’s actually there, which is a big part of why rendering is so much faster and more practical for real-time applications.
Volinga Hardware, Workflow & Integration
DP: So, what kind of machine do I need to have this running?
Fernando Rivas Manzaneque: It depends on how long you’re willing to wait. Our latest release will run on your A2000 without breaking a sweat. An RTX 3080 works fine too. The catch is time. Take that 2,000-image example at 2K resolution and you’re looking at roughly a day and a half of processing. Personally, I think that’s completely reasonable. You set it running, you go do other things, and the next morning you have a splat. It’s not instant, but it’s not asking you to babysit it either.
If you want to push things faster, our most frequent users tend to work on an A6000 or an ADA A6000. That’s also the card that tends to show up in virtual production setups, so there’s often no additional hardware investment if the studio is already equipped for that pipeline. The hard minimum is an Nvidia RTX 3000 series or above, and yes, that does mean Nvidia specifically. No way around that one for now.
On the cloud question and the security concerns: everything in Volinga processes locally, on your machine, offline. The only moment you need an internet connection is to validate your credentials at login. And for clients with stricter requirements, we offer offline licences that remove even that dependency. No data leaves the building at any point in the pipeline. We can accommodate both ends of the spectrum, from studios with relaxed policies to those with the kind of security requirements that make IT departments very popular at meetings.
Volinga Practicalities – Cameras, Colours and Clean-Up
DP: So, let’s say I shoot this with my little point-and-shoot camera , 12 megapixels, nothing fancy. How do I make sure that the Unreal scene actually looks like the real place?
Fernando Rivas Manzaneque: Colour accuracy is something we’ve thought about carefully, because it’s where a lot of tools quietly drop the ball. The short answer to the point-and-shoot question is: yes, it works, but you won’t get the full benefit of the colour pipeline unless you’re shooting with a camera that actually has the dynamic range and colour depth to justify it. A 12-megapixel consumer JPEG will give you a perfectly usable splat. It won’t give you a wide-gamut, scene-referred asset that responds beautifully to relighting in Unreal. That requires more to begin with.
Here’s how Volinga Suite currently handles it. The pipeline has two branches, depending on what you feed it. If you provide EXR files, we assume ACES 2065-1 as the input colour space and treat the data accordingly. If you provide anything else, JPEG, PNG, or standard video frames, we assume sRGB with Gamma 2.2. It’s straightforward and honest about what it’s doing.
For productions that want the full ACES pipeline, the workflow is to take your camera’s RAW footage, convert it to ACES 2065-1 EXR using a DCC like DaVinci Resolve, and feed those EXRs into Volinga Suite (There is a Netflix Technology Ressource explaining how to do that, with a charming accent and everything you need to know on how to do this: ACES Deliveries in DaVinci Resolve – relevant part starts at 7:50).
This gives you a colour-accurate, wide-gamut splat from the start. But it’s worth being honest: this only makes sense if your camera has enough dynamic range and colour gamut to survive that conversion. Running consumer camera footage through an ACES pipeline doesn’t add information that isn’t already there. It just preserves what you had more faithfully.
Where things get more interesting is in the Unreal plugin. That’s where we integrate OCIO properly. Regardless of where a splat came from, the plugin uses OCIO to convert the asset’s colour to the Working Colour Space configured by the user in Unreal Engine. That means the splat respects the project’s colour management setup rather than just landing as a raw texture and hoping for the best.
On ACES 2065 being updated: you’re not misremembering. ACES 2.0 has been released by AMPAS. The 2065-1 colour space itself remains unchanged, so the interchange format is stable. What ACES 2.0 replaces is the Output Transform pipeline, the classic RRT plus ODT combination, with a new unified transform that handles highlights more physically and performs considerably better in HDR contexts. For Volinga’s purposes, the core input handling stays the same. The implications are more on the display and grading side of the pipeline, but it’s good news for anyone building workflows that need to stay consistent across SDR and HDR deliverables.
DP: So, I bring this stuff together, Volinga does its thing, how do I clean the files up? Can I tweak them inside the Suite or is this done in the Unreal Plugin?
Fernando Rivas Manzaneque: The Volinga Suite includes a built-in editor, and it’s worth spending a moment on what it actually does, because it’s where the difference between a raw splat and a production-ready asset gets made.
The primary purpose is cleanup. Every capture picks up things you didn’t want: floaters, artefacts, background elements, and people who walked through the frame at the wrong moment. The editor lets you identify and remove those without touching the rest of the scene.
Beyond cleanup, you can combine captures. If you shot a space in multiple sessions or captured different environments that need to live together, you can merge them into a single asset inside the Suite. The reverse is also possible: if a single capture contains multiple elements that need to behave independently later in Unreal, you can isolate them as separate assets right here, so each one can be moved, hidden, or manipulated individually once you’re in the engine.
The toolset is deliberately varied because what works best depends entirely on what you’re trying to select and how the Gaussians are distributed in that area. You have brush tools, lasso selection, and volumetric selection using shapes like spheres and cubes. You can also select individual Gaussians directly when you need surgical precision.
There are two visualisation modes to support this. The standard randomised view shows the splat as you’d normally see it. The second mode displays Gaussian centroids and renders each Gaussian as a ring, which makes it much easier to see exactly where one Gaussian ends and the next begins. When you’re trying to make a precise selection in a dense or complex area, that distinction matters a lot.
Once you move into Unreal, the editing options extend further. You can use what we call Cut Volumes to mask out regions of a splat’s content directly in the engine, without going back to the Suite. We also have Proxies, which allow you to edit Gaussian material properties to a certain degree. It’s not full material authoring, but it gives artists meaningful control over how the splat responds to the scene around it.
If you want to see the editor in action, we’ve put together a walkthrough: Volinga Suite Editor.
DP: So, I added a dinosaur into my room, but it looks totally out of place. Shocking, I know. Can we do something about the lights?
Fernando Rivas Manzaneque: The short answer to the dinosaur looking out of place: yes, we can do something about the lights, and it’s more capable than most people expect.
The Volinga plugin is designed to make Gaussian splats behave like any other asset in Unreal. You add a light source, and the splat responds to it. Change the colour, change the intensity, mix multiple lights: the gaussians react, and everything in the scene stays coherent. This is what makes relighting genuinely useful for virtual production rather than just being a demo feature.
We currently support two relighting methods. The first estimates normals directly from the Gaussian data itself. It works well, requires no additional work from the artist, and gets you most of the way there quickly. The second method lets the user provide a Proxy Mesh, which we use to derive more accurate normals. This produces higher-quality results and gives you more control, but it does require some manual work. That said, it’s nowhere close to the kind of manual effort that traditional photogrammetry cleanup demands.
Now, the shadow question. You spotted a real limitation: Gaussians can receive shadows from other 3D objects in the scene, but they don’t cast shadows onto other elements. It’s one of the current constraints of the approach. The workaround, if you want proper shadow casting, is the Proxy Mesh again. Because Unreal treats the mesh as geometry, it casts shadows correctly, and the visual result reads as expected.
If a Proxy Mesh isn’t an option, whether for lack of time, resource, or skill, contact shadows are a practical alternative. They carry a small performance cost, but they get you believable grounding and interaction between your splat and the rest of the scene without any geometry work.
On-scene complexity and multiple lights: Unreal is extremely good at managing large numbers of light sources, and you benefit from all of that directly. We haven’t reinvented anything here. What we’ve done is make Unreal understand and illuminate Gaussians as a first-class element. Once that translation is in place, everything Unreal can do with lighting, you can do with your splats.
DP: So I just need a camera, the Volinga software, and Unreal Engine, that’s it? Can I use any other 3D tools with Volinga?
Fernando Rivas Manzaneque: A regular camera is genuinely enough to get started, and the previous answers cover how to get the most out of one. But there are other routes into the pipeline, and some of them change the rules of the game considerably.
Our partner Xgrids makes capture hardware specifically designed to reduce the constraints we’ve been talking about. Their devices combine a LiDAR sensor with an IMU, an “inertial measurement unit”, which means that while you’re moving through a space, the device is simultaneously solving its own position in 3D. Structure from Motion is happening in real time during capture, rather than being solved retroactively from images alone. The practical effect is that the strict requirements around parallax and pattern overlap become much more relaxed. They’re still good practices, but they’re no longer make-or-break.
Xgrids devices also come with a companion mobile app that gives you live feedback during capture, showing which areas of the space have been covered and which are still missing. That real-time map changes the capture experience significantly: instead of hoping you got everything and finding out during processing, you can see the gaps and fill them on the spot.
The main limitation with these sensors is on the imaging side. The cameras are 8-bit, and resolution isn’t their strongest point. For most use cases you can work around this with their HD Enhancement feature, Grids Probe, or simply by investing more time in the capture. The output is fully compatible with the Volinga Unreal plugin, so the pipeline stays exactly the same regardless of which hardware you used.
As for which Xgrids device makes sense: it depends heavily on what you’re capturing. The PortalCam is excellent value and particularly well-suited to interiors. For large outdoor environments you’ll cover more ground and may need to swap batteries mid-session, but it handles it. The XLK2 trades a slightly weaker image sensor for a longer-range LiDAR, which means fewer passes needed to cover a large space. If you’re working on expansive environments, that trade-off often makes sense.
On 360 cameras, Insta360 and similar devices, as well as certain drone setups, are producing genuinely promising results. The limitation there is sensor size relative to the field of view: you’re covering a lot of angles at once, but the detail in any given direction suffers for it. The practical fix is to add closer, more targeted passes over the areas that need finer detail. For a quick capture or a previz-quality environment, 360 solutions are hard to beat. For a hero set, they work best as a starting point that you refine.
Production Realities: Capture, Errors & Unreal Integration
DP: So, let’s say I had a big run-around in the room with the PortalCam. I did that, imported it into Volinga, it did its thing, I cleaned it up, now I have a perfect 3D version of my room. Hopefully I didn’t move anything during shooting. How much space for errors is there in the data?
Fernando Rivas Manzaneque: Film sets are chaotic by nature, and Gaussian Splatting is not entirely forgiving of that chaos. But there’s more room to manoeuvre than you might think, and the tools exist to fix most of what goes wrong.
The good news first: Structure from Motion is quite resilient to occasional bad frames. If a person walks through the background in three shots out of two hundred, the algorithm has enough consistent data from the rest to effectively outvote those frames. The reconstruction will handle it. The same logic applies to minor lighting shifts. A handful of inconsistent images in a large dataset will usually just produce a small amount of noise or a faint floater in that area, rather than breaking the whole reconstruction.
Where things get harder is when the problem is systematic rather than occasional. If lighting changed significantly halfway through a capture session, if an object was in one position for the first half and a different position for the second, or if a large area of the set was consistently obstructed, the model has genuinely contradictory information and will invent geometry to reconcile it. That’s where you get the floaters we mentioned earlier, and no amount of processing will resolve a fundamental inconsistency in the source data.
The Volinga Suite editor is your main tool for dealing with what slips through. Artefacts, floaters, ghost geometry from people or objects that moved: these can all be identified and removed in the editor without touching the rest of the scene. For a film set context, a good workflow is to treat the first pass through the editor as a deliberate cleanup step, not an optional one.
What the editor cannot do is invent information that wasn’t captured. If you missed an angle, that gap is real and it will show. The solution there is to go back and capture it. With devices like the Xgrids PortalCam and its companion app, you can catch those gaps during the session itself, before you’re back in the office and the set has been struck. That real-time coverage map is genuinely useful on a production where time on set is limited.
The practical advice for anyone working in a live production environment: lock down what you can, capture with enough overlap that individual problem frames don’t matter, and use the editor to clean what you couldn’t control. The system is more forgiving than old photogrammetry pipelines, but it still rewards discipline during capture.
DP: So let’s say the pipeline works, everybody knows what to do, and somebody taped the DOP and the gaffer into their corners so they don’t move around the set. In your experience, how long should I plan to scan or capture the set or location?
Fernando Rivas Manzaneque: Honestly, this is one of the hardest questions to answer straight, because it really does depend.
The most instructive example I can point to is the work the CBS VFX team did with us on productions including The Neighbors, Frasier, and NCIS. They built a custom five-camera rig using Nikon D850s, which effectively let them capture five angles simultaneously instead of one. With that setup, they were consistently getting episodic TV sets done in around ten to fifteen minutes of capture time. That’s a room-sized set, production-ready quality, taped and out of the way before the crew notices you were there. You can read the full breakdown of how that pipeline came together on our blog.
And here is an Example from the Frasier production. Guess which side is the splat!
For a solo operator with a PortalCam, you can move fairly quickly. A single interior space, covered with enough overlap for a clean previz or a solid reference environment, is realistically fifteen to twenty minutes of walking. If that’s all you need, you’re done before lunch.
The numbers start climbing when the space gets larger or when you’re going for hero-set quality. A detailed, complex interior with fine geometry, clean edges, and no gaps to clean up in post can take two to three hours of capture, sometimes more. Large outdoor environments can go further still. There’s no dishonest way to say it: the quality ceiling is high, and reaching it takes time on set. How much time it takes depends on the space, the target quality, and, honestly, how much care the person with the camera is willing to put in.
The encouraging part is that the entry point is low. You can have something genuinely useful in under half an hour. Where you go from there is a question of ambition and available time, which on a production is usually the same thing
DP: And when the splats get serious, can Volinga handle parallel processing?
Fernando Rivas Manzaneque: Parallel processing is possible, with one important clarification on what that actually means in practice. You can absolutely run multiple splat training jobs simultaneously across different machines. What you cannot do is split a single splat and train it on multiple machines simultaneously. Each job runs on one machine from start to finish.
So if you have a large environment that needs to be broken into sections, the workflow is to do that division upstream, using a tool like RealityCapture to split the scene, and then send each sub-splat to a different machine. Different operators can manage different sections in parallel, and the resulting assets can be combined later in the Volinga Suite editor. It adds a step, but it scales.
For studios managing multiple projects or locations at once, Volinga Suite includes a Job Queue Manager. You line up everything you need to process, launch the queue, and walk away. The machine works through the list overnight or over a weekend without anyone needing to babysit it. For a team running a busy shoot schedule, that’s a meaningful quality-of-life feature.
At the Enterprise level, we offer more scalable deployment options that can integrate with a studio’s existing render farm infrastructure. The specifics depend on the setup, but the goal is to fit into the pipeline rather than sit next to it.
On the question of tiers more broadly: the Personal plan is focused on the Unreal plugin. If you’re an independent artist who already has splats and wants to bring them into Unreal, that’s the entry point. The Suite, with its full processing and editing capabilities, comes in at the Studio level and above.
And yes, we’re working on an automatic end-of-day shutdown feature for public broadcasters. Legal told us we might have to call it something other than “go home, it’s seven o’clock,” but the functionality is basically the same.
DP: How big are these splat files anyway?
Fernando Rivas Manzaneque: File size in Gaussian Splatting is primarily a function of the level of detail you’re capturing, not the physical volume of the space. That distinction matters, and a couple of examples make it concrete.
We have an auditorium in our sample projects that weighs around 800 MB. It seats three to four hundred people, includes a full orchestra setup on stage, and has solid detail throughout, though it’s optimised as a background environment rather than something you’d push into close-up. Contrast that with an abandoned workshop, roughly 40 square metres, which comes in at around 2 GB. It’s a fraction of the physical size, but it’s dense with fine detail across every surface.
Up to about 5 GB, our plugin handles assets comfortably. Once you start capturing larger environments, full city blocks, entire backlots, files can reach 10 to 15 GB, and that’s where things currently get harder. Real-time performance starts to degrade, and we’re actively working on improving how the plugin handles those larger assets. It’s one of the areas where the technology is still catching up with what productions want to throw at it.
DP: So, now the splat is fine. What do I do with it?
Fernando Rivas Manzaneque: Once the Volinga Suite has finished processing, you export the result as either a PLY or our native NVOL format, which carries additional metadata useful downstream.
From there, the path into Unreal Engine is about as straightforward as it gets: you install the Volinga plugin, drop the file into your content browser, and drag it into the scene. At that point, it behaves like any other Unreal asset. You can position it, relight it, combine it with other geometry, and send it out through whatever Unreal-based VP pipeline you’re using.
DP: Why is that a plugin, actually?
Fernando Rivas Manzaneque: Unreal Engine doesn’t natively understand Gaussian Splats. Despite all the noise the technology has been making, and there’s been a lot of it, there’s no built-in 3DGS support in the engine yet.
There are approaches that attempt to represent splats using Niagara particle systems inside Unreal, but they’re significantly limited: relighting is constrained, file sizes become unwieldy, and the visual quality suffers at scale. Trying to bake splats into meshes produces noisy, artefact-heavy geometry that doesn’t hold up in production.
Our approach is different. We render in parallel alongside Unreal’s own rendering pipeline, using texture sharing to pass the splat data into the engine without going through geometry. This means we preserve the full visual quality of the Gaussian Splat while retaining almost all the benefits of working inside Unreal: the lighting system, the compositing, the real-time performance. It’s less of a conversion and more of a handshake between two renderers running side by side.
DP: So if the plugin just takes the splat file and hands it off to the engine, could I use splats that I made in PostShot, RealityCapture, or some other tool?
Fernando Rivas Manzaneque: Yes, and that’s very much by design. The goal has always been an open ecosystem, not a closed one. The Volinga Unreal plugin will load PLY files from other software, provided the vendor’s own licensing allows export in that format. RealityCapture, for instance, doesn’t generate Gaussian Splats at all, so that’s not a pipeline question. But tools like PostShot or other 3DGS training software that export standard PLYs work fine with the plugin. You can capture and train your assets wherever makes sense for your workflow and bring them into Unreal through Volinga.
DP: Let’s say Unreal finally understands my splat. I’ve got my virtual stage, and I’m blasting it out through LiveFX or whatever flavour of Unreal-based VP tool we’re using today. Now something changes, as it always does. Can I round-trip that data back to Volinga?
Fernando Rivas Manzaneque: The honest answer here is: some things are fast, and some things hit a wall. Relighting is the easy one. As we covered, the Volinga plugin integrates splats into Unreal’s lighting system, so if the brief changes and you need to shift the look of the environment, you work with Unreal’s lights directly. No round-trip required, no re-export, no waiting. That part is genuinely fast enough for on-set iteration.
Geometry cleanup is also manageable. If something needs to be removed or adjusted, you go back to the Suite, make the edit, re-export, and reload in Unreal. It’s not instant, but the pipeline is short enough that it’s a realistic option between setups rather than something that derails the day.
Where you hit the fundamental limitation of the technology is texture editability. A Gaussian Splat bakes the appearance of a scene into the Gaussians themselves. There’s no texture map to swap out, no material layer to override in the traditional sense. What the camera saw is what you have. You can’t repaint a wall or change a logo on a prop after the fact, the way you might with a photogrammetry mesh and a texture atlas.
The workaround, and it does work, is the ProxyMesh approach we mentioned in the relighting section. By associating a proxy mesh with a splat, you get access to material overrides that can drive changes in how the surface responds to light and how it reads visually. But this requires pre-production planning.
You need to have built the proxy and set up the material relationships before you’re standing on the brain bar at two in the morning wondering why the set looks wrong. If that groundwork is in place, you have meaningful flexibility. If it isn’t, your options are more limited.
The practical takeaway for a VP supervisor is: plan for the splat’s editability constraints the same way you’d plan for anything else that can’t be changed on the day. Relighting is yours to play with freely. Anything more structural needs a workflow decision made before the shoot, not during it.
For the Curious and the Budget-Conscious
DP: So, I have my pictures, how do I get the Volinga Suite?
Fernando Rivas Manzaneque: There’s a free version for non-commercial use, which is the right place to start. It includes both the Suite and the plugin, with a watermark on the output. For anyone doing R&D, testing the pipeline, or just trying to understand whether the technology fits their workflow, it gives you full access to the toolset without any financial commitment.
For learning materials, we have our documentation at docs.volinga.ai, which covers the full pipeline from capture to Unreal. And our blog at web.volinga.ai/use-cases is where we put case studies, workflow breakdowns, and practical how-tos from real productions. If you want to understand how other teams have integrated this into their pipelines before building your own, that’s the place to start. If you want to have an Unreal Scene to play around with: We have a few samples for free, on our website HERE.
DP: Let’s say you’ve convinced me. How much of my budget should I set aside for the Volinga part of the pipeline?
Fernando Rivas Manzaneque: Our strong preference is that people try everything end-to-end first, with the free non-commercial version (Create a free account, downmload all the parts, plus demo material). The full pipeline, Suite to plugin to Unreal, is available without cost for R&D purposes. That’s intentional. We’d rather someone spend a few weeks genuinely understanding how Volinga fits into their workflow before they make any purchasing decision. By the time they come to us commercially, they know exactly which tools they need and why. If you need a single-page-thing for your C-Suite: here is a ultra-short introduction as a PDF.
On the numbers: it depends heavily on how many seats are required and which combination of tools makes sense for the studio. As a starting point, a single Suite licence plus plugin is in the region of €2,000 per year. For studios scaling up across multiple operators or integrating into larger production pipelines, the conversation moves to the Studio or Enterprise tiers, and at that point, it’s worth talking to us directly about what the pipeline actually looks like.
There are a few different versions: The Volinga Plugin has PLY Rendering in UE, Relighting, Post-Process and DoF, Multiple 3DGS Models and Executable Project Shipping.
The “Plugin Pro” has all of that, plus CULL Volumes, HDR and ACES support, nDisplay Support, Color Correction Region Support, Mesh Support and talks to RenderStream, Aximmetry, and LiveFX.
And the Suite for Studios and Enterprise does HDR and ACES pipeline support, 4K and 8K processing, has the Job Queue Manager as well as Reality Scan and Metashape camera registrations imports. And yes, you can try all of that for free.
Scaling Up, Surviving Production & the Occasional Cat Scan
DP: So, what happens when we go big? Let’s say we’re not talking about a room anymore but a full set, a back-lot, or an outdoor environment. How well does Volinga scale?
Fernando Rivas Manzaneque: The performance envelope we’re most comfortable with right now is up to around 5 GB on a well-equipped workstation, an A6000 or equivalent. Within that range, the plugin handles assets without meaningful degradation. The auditorium and workshop examples from earlier both live comfortably in that space.
Where things get harder is when you start capturing environments at the city-block scale. Those assets can reach 10 to 15 GB, and at that size, the real-time performance starts to feel the weight. It’s not a wall, but it’s not effortless either. We’re actively working on it. Level-of-detail (LOD) systems and spatial partitioning are both on the roadmap, specifically to address this, so the ceiling will move.
But right now, if you’re planning a production around very large outdoor environments, it’s worth having that conversation with us before you’re standing in front of an LED wall, wondering why the frame rate dropped.
DP: What happens if I want to bring my splats into another DCC, say, Blender, Houdini, or Nuke for compositing or lighting?
Fernando Rivas Manzaneque: The short answer is that the ecosystem is moving fast, and the PLY format is doing a lot of the heavy lifting. Blender already has plugins that handle PLY import, so bringing splats into a Blender scene for compositing or reference is feasible today. For Houdini, the GS Ops plugin is genuinely good and worth considering if your pipeline runs there. Nuke has had PLY-based plugins for a while, and more recently, Nuke added native Gaussian Splat support through USD, which is a significant step.
On our side, USD integration is on the Volinga roadmap. The goal is to make sure that what you build and process in the Suite can flow directly into Nuke and other USD-capable tools without additional conversion steps. We see Unreal as the primary rendering destination for now, but the intention has always been to stay interoperable rather than become a closed system.
DP: So, what’s the production-readiness verdict? If I were a VFX supervisor trying to convince my producer this will actually work, what would you tell me?
Fernando Rivas Manzaneque: My honest verdict: production-ready, particularly for virtual production. Not experimental, not a proof of concept you’d bring to a shoot and hope for the best. Production-ready.
The most significant project we can point to right now is Netflix’s Berlin and the Lady with an Ermine, produced by Vancouver Media, the team behind Money Heist. The production needed to shoot a nighttime sequence along Madrid’s Calle Alcalá, combining large-scale stunt work with intimate close-ups of two characters. The close-ups were shot on a virtual production stage with Gaussian Splats of the real location as the backdrop, and the final edit is seamless. The audience cannot tell which shots were on location and which were on the LED wall. You can read the full breakdown here, and if you want to see it for yourself, it’s on Netflix now, Episode 1, minute 32:20.
Beyond that, we’ve worked with CBS VFX across multiple episodic productions, with Envidio, and with a number of other productions that haven’t been announced yet. The headaches we’ve hit along the way are mostly the ones we’ve already talked about: capture discipline, colour pipeline setup, and the editability limitations of splats. None of them are dealbreakers once you know they’re coming. The teams that plan for them consistently get results they’re proud of.
DP: You’ve got to have at least one story of when something went completely off the rails. What’s the weirdest or most chaotic capture session you’ve seen someone attempt with Volinga?
Fernando Rivas Manzaneque: There’s one that comes up more often than you’d expect, and it’s become something of an informal rite of passage for anyone new to the workflow.
Gaussian Splats capture everything the camera sees. That includes mirrors. And reflections in mirrors are, by definition, pointing back at the person holding the camera. We’ve had more than one capture session where the artist doing the walkthrough forgot to account for this, walked through a bathroom, a dressing room, or any space with a well-placed mirror, and later discovered that they had very thoroughly documented themselves in a state that was not quite production-appropriate. The splat doesn’t judge. It just captures.
The lesson, now standard in any capture briefing we give: treat every reflective surface in the space as a camera pointed directly at you. Dress accordingly. We’ll leave it there.
DP: And finally, what’s next for you and for Volinga?
Fernando Rivas Manzaneque: A few things we can talk about, and a few we can’t yet. On the research side, the field is moving extremely fast and we and our partners intend to stay at the front of it. One of the directions we’re most excited about is 4D Gaussian Splatting, which extends the model into the time dimension and opens up possibilities for capturing dynamic scenes rather than static ones. Getting that into Unreal is something we’re working towards.
Scalability is the other major engineering focus. We’re building our own level-of-detail system and spatial partitioning specifically to overcome current size limitations. The goal is to make large outdoor environments as manageable as interior sets are today.
And on the Suite side, the direction is towards more AI-driven workflows that give artists greater control and editability over their splats. The fundamental constraint of splats right now is that they bake the appearance at capture time and resist change afterwards. We want to change that, gradually but meaningfully, through tools that feel native to a production workflow rather than requiring a research background to operate.
Follow us on LinkedIn, and get a heads up on what happens!
https://www.linkedin.com/company/volinga/posts/
