Gaussian Splatting Stops Being a Demo: Production Pipelines in 2026

In August 2023, the Gaussian Splatting paper hit arxiv and 3D Twitter lost its mind. The demo was a NeRF-quality scene reconstructed from a phone video, rendered at 100+ fps on a laptop GPU. It looked like cheating.

Three years later, splats are no longer a demo. Postshot is in mainstream archviz workflows. Luma AI's Genie is feeding cinematic backdrops to indie film teams. Polycam ships splat capture as a one-tap mobile feature. Unreal Engine 5 has community plugins for real-time splat playback. The technique has crossed the gap from research curiosity to production tool — but the pipeline is awkward, the failure modes are real, and most teams adopting it are quietly figuring out where it actually fits.

This is what we have learned about that fit.

What Splats Actually Are (One Paragraph)

A Gaussian Splat scene is a cloud of millions of soft 3D ellipsoids, each with a color, opacity, and view-dependent shading direction. Instead of meshes and textures, you train this cloud from a set of images so that, when projected back to those camera positions, it reproduces the original views. Rendering is rasterization — fast — not ray marching like NeRFs. That is why splats render in real-time and NeRFs largely do not.

Why 2024-2025 Was the Inflection

Two things changed:

Training time collapsed. The original paper trained a scene in 30 minutes on an A100. By mid-2025, consumer-tool training was down to 2-5 minutes on a 4090 for a typical 200-image capture. Phone-side training (Polycam, Scaniverse) appeared, doing it in under a minute on-device for small captures.

Renderers shipped to engines. Native splat rendering arrived in Unreal Engine 5 (multiple plugins), Unity (the Aras Pranckevičius implementation became canonical), Three.js (Mark Kellogg's renderer), Blender (KIRI Engine + native plugin), and Spectre as a standalone DCC. Splats stopped being trapped in the training tool.

Together: capture in the morning, render in your game engine the same afternoon. That changed everything.

Where Splats Win in Production

Three real categories where studios have shipped splats in 2026:

Cinematic backdrops

Indie virtual production teams use splats for distant background plates: a captured forest, an alley, a cityscape. Splats render fast enough to be a real-time backdrop on an LED volume, which traditional photogrammetry meshes never managed (mesh complexity exploded for similar visual fidelity). The trick: splats are read-only — you cannot easily place a chair on a splat floor and have it occlude or shadow correctly. So they live behind the action, not within it.

Architectural and real-estate walkthroughs

Capture a property in 30 minutes, train overnight, ship a Three.js viewer the next day. The viewer renders smoothly on a phone. This use case has crushed traditional photogrammetry for high-end real estate listings — splats handle reflective surfaces (windows, polished floors) far better than triangulated meshes do.

Game backdrops and "skybox 2.0"

Some game studios have started using splats as next-gen skyboxes — distant city scenes, mountainscapes, hangar interiors viewed through a window. The splat occupies the far end of the frustum where you would once have used a matte painting or a low-poly mesh. Performance is acceptable because the splat is far enough that LOD and culling work well.

Where Splats Still Lose

The honest list. Anyone selling splats as a meshes-killer is overselling.

Animation and deformation. Splats are static point clouds. There is no "rig a splat" workflow. Recent papers (Dynamic Gaussian Splatting, 4D Splats) show animation works in research but not in any production tool yet. If your asset has to move, you reach for a mesh.

Editing. You can mask, crop, and clean up splats in Postshot or Spectre. You cannot easily add geometry, change topology, or paint a splat the way you paint a mesh. The output is closer to a photo than to a model.

Sharp specular at glancing angles. Polished metal and glass viewed almost edge-on still look smeared. The view-dependent shading per splat smooths out the specular highlight that real surfaces produce. Some 2025 papers improved this; it is not yet in mainstream tools.

Engine compatibility. Splat plugins exist for UE5 and Unity but are not first-class citizens. Lumen and Nanite do not interact with splats. Lighting baked into the splat at capture time does not respond to dynamic engine lights. So your splat scene looks great at the time of day you captured — and only at that time of day.

Storage. A high-quality splat scene is 200-800 MB. A photogrammetric mesh of the same scene compresses to 50-100 MB. Splats are not edge-friendly yet; streaming-splat formats (PLY-based, KSplat, the new SOG draft format) are immature.

The 2026 Production Pipeline

What teams shipping splats today actually do:

1. Capture. 100-300 photos or a 1-2 minute video. Phone is fine for static scenes; mirrorless body for hero captures. Diffuse light beats hard shadows.
2. Train. Postshot (paid, the production default), Luma AI Genie (cloud, fast), Polycam (mobile-first), or Nerfstudio (open source, research-flavor). Training writes a .ply or .splat file.
3. Clean up. Mask out floaters and background junk. Postshot and Spectre have decent tools. Expect to spend 30-60 minutes per hero scene.
4. Convert if needed. Game engines often want a specific binary format. KSplat, SOG, or vendor-specific formats. The lossless PLY tends to be too big to ship.
5. Integrate. Drop into UE5/Unity via the splat plugin, position relative to your gameplay camera, and treat it as a static set piece. Do not put gameplay objects intersecting the splat.

The Tools Landscape

Where things sit in 2026 (incomplete, but a useful map):

• Postshot — current production default. Windows, paid. Best balance of quality, training speed, and cleanup tools.
• Luma AI Genie / Luma Capture — cloud-based, the easiest path. Good when you do not want to manage a GPU.
• Polycam / Scaniverse — mobile-first, on-device or cloud. Best for archviz and casual users.
• Spectre — DCC-style splat editor. Less mature than Postshot but improving fast.
• Nerfstudio (with Splatfacto) — open source, research-grade. Where new techniques land first.
• KIRI Engine — mobile + Blender plugin path; covers capture and integration.
• Unity / Unreal community plugins — runtime renderers. Aras Pranckevičius for Unity, several for UE5.

What This Means for 3D Artists

Splats do not replace your meshing skills, your topology eye, or your sense of what makes a scene readable. They add a new tool for a specific category of asset: the high-fidelity static set piece you would once have spent days building from scans + retopology + texturing.

The 2026 reality: most production scenes are now hybrid. Hero gameplay assets stay as meshes (you need to rig, animate, and modify them). Static backdrops and reference captures move toward splats. Both pipelines coexist in the same scene file — the engine renders splats and meshes side by side, with the splats acting as the unmoving stage and the meshes acting on it.

The killer application — animated, lightable, fully editable splats — is still in research. When it ships to mainstream tools, the photogrammetry-mesh pipeline will be in the same place that the spline-modelling-only pipeline was in 2008: still in use, but no longer the obvious default.