GaussianSplat Gem for Open 3D Engine (O3DE)

A gem for O3DE that renders 3D Gaussian Splatting scenes. Each splat is a 3D Gaussian defined by a position, anisotropic scale + rotation (covariance), opacity, and a view-dependent color encoded as spherical-harmonics (SH) coefficients. Splats are imported from 3DGS-standard .ply files and rendered through a tile-based compute rasterizer.

A bit of context

• Open 3D Engine - an open-source game & simulation engine. O3DE is extendable through modules called Gems. This is one such Gem.
• 3D Gaussian Splatting - rasterization-based technique used in the field of real-time radiance fields for representing and rendering photorealistic three-dimensional scenes from a sparse set of two-dimensional images.

Requirements

The gem targets Linux with the Vulkan RHI only, and has not been tested on any other platform. Other platforms are disabled: their CMake returns early and builds no targets. Supporting them should be possible, but may take some work.

SM6.0 wave intrinsics (RHI::DeviceFeatures::m_waveOperation), guaranteed by the Vulkan 1.1 target the shaders set. There is no non-wave fallback path - see Docs/Rasterizer.md's "Wave intrinsics are a hard requirement".

The gem was tested with O3DE 2605.0 and 2610.0.

Examples

Below are some examples of scenes rendered with GaussianSplat gem in O3DE:

Screenshot	Attribution
	"Cluster Fly" by danylyon, licensed under CC BY 4.0.
	"Winter Garden" by andriishramko, licensed under CC BY 4.0.
	"International Space Station - Synthetic scene" by tmate, licensed under CC BY 4.0.

Usage

1. Place a Gaussian Splat .ply file inside the project's Assets folder.
2. Create an entity with a Gaussian Splat Component.
   
3. Select Gaussian Splat asset in the selector.
   
4. Adjust parameters to suit your needs.
   

Runtime tunables (GaussianSplatComponentConfig)

Editor field	Purpose
Gaussian Splat Asset	The .gsplat asset reference to render
Opacity Multiplier	Global alpha multiplier applied on top of each splat's stored opacity; [0, 1], so it can only attenuate
Scale Multiplier	Global size multiplier applied on top of each splat's stored scale; [0, 10]. Resizes each splat in place, unlike the entity Transform's scale, which also changes splat spacing
Max SH Degree	Clamps SH evaluation below what the asset stores - 0 renders flat (DC-only) color, useful as a quality/perf knob. Also shrinks the GPU SH buffer

Global Configuration

Radix-sort entry capacity

The radix-sort entry buffers are sized to a fixed capacity, set once when the feature processor activates, from the Settings Registry key /Robotec/GaussianSplat/RadixEntryCapacity (default 16 * 1024 * 1024), defaulted in Registry/gaussiansplat_settings.setreg.

A value of 0 is treated as invalid and falls back to the compiled-in default. The capacity drives the two largest allocations in the gem: at the default, each of the two ping-pong entry buffers is ~192 MB. If those allocations fail, the pass chain is left out - and retried the next frame - rather than taking the pipeline down; lower this value if the splats never appear.

Depth output

The splats' own depth is merged into the render pipeline's linear depth buffer (DepthPrePass's DepthLinear) by a composite pass, the depth-side counterpart of the color composite. The key below is read once at activation, from Registry/gaussiansplat_settings.setreg:

• /Robotec/GaussianSplat/DepthCoverageThreshold (default 0.02) selects which splat a pixel's depth comes from - the first one, front to back, whose compositing pushes the pixel's accumulated coverage to this value. Near 0 this is the nearest splat that contributes at all; 0.5 gives the median depth. Values outside (0, 1] fall back to the default.

Because the splat pass chain runs after PostProcessPass, the merged depth is visible to consumers placed after the chain (addressable as GaussianSplatParent's DepthLinearInputOutput), not to the engine's own post-processing, which reads depth earlier in the frame. See Docs/Rasterizer.md's "Depth output".

Data flow

The pipeline is GPU-only from end to end: per-splat screen-space projection, culling and SH color evaluation run in a compute pass, then a chain of compute passes expands each splat into per-tile entries and radix-sorts them by a combined (tileIndex, depth) key, before a tile-shading compute pass composites each tile's sorted splats front-to-back. The sort is a six-round, 8-bit LSD Onesweep radix sort with decoupled look-back, written from scratch for O3DE.

See Docs/Architecture.md for the design-decision record (asset format, coordinate conversion, GPU buffer layout), Docs/Rasterizer.md for the rendering pipeline specifically.

---
config:
  flowchart:
    curve: linear
    htmlLabels: true
---
flowchart TD
    PLY["3DGS-standard .ply\n(position, scale, rotation,\nopacity, f_dc, f_rest)"]
    Builder["GaussianSplatBuilder\n(content-sniffs vertex properties,\napplies sigmoid/exp/normalize,\nY-up -> Z-up basis change)"]
    Asset[".gsplat asset\n(GaussianSplatAsset)"]
    Controller["GaussianSplatComponentController\n(shared by Client + Editor components)"]
    FP["GaussianSplatFeatureProcessor\n(owns per-pipeline GPU buffers,\ninjects + sizes the pass chain)"]
    CullGroup["GaussianSplatProjectCullGroupPass\n(VisibleCounterClear + one ProjectCull\npass per cloud; compute: projection,\nculling, SH color, tile-touch ranges)"]
    BuildKeys["PrepareBuildKeysDispatch -> RadixBuildKeys\n(compute: expand each splat into\nper-touched-tile (tileId, depth) keys)"]
    Prep["RadixPrepareSortDispatch\n(clamped entry count +\nindirect dispatch args)"]
    Hist["GlobalHistogram -> DigitBase\n(all 6 rounds' histograms in one\npass over the entries)"]
    Sort["6x GaussianSplatRadixOnesweepPass\n(one dispatch per round;\ncompute: fused histogram/look-back\nscan/stable scatter per round)"]
    Unpack["RadixUnpack\n(compute: derive per-tile\n(offset, count) ranges)"]
    Shading["GaussianSplatTileShadingPass\n(compute: per-pixel Gaussian falloff,\nfront-to-back alpha compositing,\nopaque-depth occlusion)"]
    Composite["GaussianSplatComposite.pass\n(fullscreen triangle: blend onto\nmain color target)"]
    DepthComposite["GaussianSplatDepthComposite.pass\n(fullscreen triangle: min-blend splat\ndepth onto linear depth target)"]

    PLY --> Builder --> Asset --> Controller --> FP
    FP --> CullGroup --> BuildKeys --> Prep --> Hist --> Sort --> Unpack --> Shading --> Composite
    Shading --> DepthComposite

Loading

Known limitations

• Very dense views can drop splats once a frame outgrows the radix-sort entry capacity. If splats vanish up close, raise /Robotec/GaussianSplat/RadixEntryCapacity (see Radix-sort entry capacity).
• Render outputs above ~16.7 megapixels (65535 16x16-pixel tiles, e.g. beyond roughly 5460x3070) are not supported - splats sort incorrectly and a warning is logged.
• Resizing a viewport or loading a cloud causes a brief hitch while the GPU buffers are rebuilt.
• Only the 3DGS-standard .ply layout is imported - there is no .splat (antimatter15 compact format) import path.

See Docs/Rasterizer.md's "Known open risks" for implementation-level caveats.

License

This gem is licensed under the Apache License, Version 2.0 (SPDX-License-Identifier: Apache-2.0).

The example screenshots in Docs/Images and preview.png (a copy of the "Cluster Fly" screenshot) are renders of third-party scenes, adapted under the licenses listed in Examples. They are not covered by the gem's license.