Visual Technology Research • Spring 2026
We assembled the team this week and immediately split into research territories. No slow ramp-up, no endless planning meetings. Everyone came prepared with clear technical areas they wanted to investigate.
The goal: understand what's technically possible before we commit to any specific project direction. Research first, decide later.
My research spans four interconnected areas that we think might work together:
These aren't separate research tracks - they're pieces of the same puzzle. If GS can't export to USD, can TD act as a bridge? If we can't manipulate splats in Unreal, can GenAI tools help us process the data differently?
Gaussian Splatting has exploded over the past year as a capture technology. You can scan real-world environments with unprecedented speed and visual fidelity. But there's a fundamental workflow problem:
GS is treated as an end product, not source material. We can't iterate.
What We Need:
Capture → Manipulate Data → Create Interactive Experiences
If we can make GS behave like editable 3D data rather than baked renders, it opens up applications in architectural visualization, virtual production, interactive installations, anywhere photogrammetry meets real-time creative control.
Quick answers before diving into details:
Got trial access to https://volinga.ai/main Friday afternoon. Installation was smooth - followed their plugin docs at https://docs.volinga.ai/ue-plugin, took maybe 20 minutes from download to first render in UE 5.6.1.
Used open-source PLY files from Sketchfab for testing - specifically chose indoor scenes with complex lighting to stress-test the relighting capabilities.
Test Hardware: RTX 4070 SUPER / Ryzen 9 3900X 12-core / 64GB RAM
Ran three specific tests to isolate performance costs:
Key Finding: The 10fps drop when enabling relighting is acceptable for real-time work. More importantly, adding extra lights doesn't scale the performance cost linearly - the system appears to batch light calculations efficiently. This suggests we could have multiple dynamic lights in a scene without destroying frame rate.
Wanted to verify that GS rendering integrates properly with Unreal's standard pipeline, not bypassing post-process stages.
Test: Bloom Effect
This confirms Volinga isn't taking shortcuts. The splats render through Unreal's proper pipeline, which means any post-process effects we need later (color grading, DOF, motion blur, etc.) should work correctly.
This is where things got interesting. Tried to export the GS scene to USD using Unreal's native USD exporter, expecting to get point cloud data or some geometric representation.
Result: Empty bounding box mesh. That's it.
Started researching Luma AI's Unreal plugin at Luma AI Documentation
Why Luma AI Might Be Different:
Their architectural approach is fundamentally different from Volinga, they render GS as Niagara particle systems rather than proprietary rasterization. This could theoretically give us:
Current Intel:
Next update: Luma AI results + TouchDesigner findings + how this is shaping our project direction. Documentation by end of week.
New Approach: LiDAR Point Cloud Plugin Documentation
It seems ply can store vertices and also triangles not unlike OBJ format.
We can modify the header and only preserve the vertex data, then format it as a XYZ, and finally use the Lidar plugin in UE to read those data LiDAR Point Cloud.
Above 110 FPS
Color data works
Relighting works
Luma.ai
TO DO: (listed in priority)
To make the USD file work correctly, the following structure is needed:
├── Single.usdc
└── obj/
└── Fern_v1/
└── Out_Growth.usd
Building interactive demos. Testing real-time performance under different conditions. Pushing the tech to see where it breaks. This is where theory meets reality.
Focusing on visual quality. Refining shaders, testing lighting scenarios, making it look good while keeping it performant. Finding that balance between beauty and speed.
AI-assisted workflow automation. Testing generative AI integration points across the pipeline. Finding where AI actually adds value versus where it's just overhead.
Real-time generative systems. Building interactive demos that respond to user input. Testing TouchDesigner connectivity with Unreal Engine.
Cross-platform asset pipeline refinement. Making sure everything works together smoothly. Optimizing performance across different systems.
Documenting everything. Writing up technical findings, successful workflows, dead ends we hit. Making it clear and shareable.
Creating a polished demo that shows what's possible. Visual presentation of research and results. Live demonstration and Q&A.
