What a Gaussian Count Actually Measures

The number needs a stage

A project containing millions of Gaussian references can still show a much smaller preview. Conversely, a modest source can become expensive when repeated throughout a scene. Both statements can be true without an error in the reported numbers. A creator needs to know whether a counter describes a request, completed stored data, placed source references, resident preview samples or the samples active for the current view. Collapsing those categories makes a capacity label look more informative than it is.

The creator-facing goal in Ashley Kalkowski’s Guassia work is to retain useful authoring detail while making the current workload understandable. This is an interface and workflow problem as well as a rendering problem. An honest count helps someone decide whether to simplify a preview, change a scene layout or preserve the original source for a later export.

Preservation and display answer different questions

Authored capacity concerns what the project can represent and retain. Display capacity concerns what the current renderer and device can process within the chosen preview settings. Shared instances introduce another distinction: repeating an asset can increase its placed count without requiring independent copies of every source record. A reduced display is therefore not necessarily a reduced saved asset.

Guassia’s September 23 notes describe explicit preview tiers alongside retained source data. The September 24 visibility work addresses off-camera instances. These controls serve different purposes. Preview selection bounds displayed work; visibility rejection avoids work outside the current camera. Neither is evidence that a project implements full source streaming, hierarchical level of detail or automatic graphics-memory eviction. The public notes explicitly preserve that boundary.

What the public browser fixture establishes

The September 23 changelog reports a browser fixture with a one-million-Gaussian source referenced five times. Its authored count was five million. The Balanced preview used 70,000 shared samples and 350,000 active samples, while preserving the source hash. The practical observation is that a bounded preview and a larger retained project can coexist. It is not a demonstration that five million full-detail visible Gaussians were drawn simultaneously.

The same entry records 84.2 milliseconds of CPU-plus-awaited-renderer work at 1214 by 758 pixels. That observation is not a GPU timestamp, a smooth 60-frame-per-second result or a transferable hardware limit. Treating it as a universal benchmark would remove the workload and measurement definitions that make it useful. No new timing experiment was conducted for this report.

A collection target is not completed geometry

The September 27 Asset Engine notes distinguish requested, committed and preview counts for large local procedural collections. Requested describes an intention; committed describes data successfully retained; preview describes a bounded view. Cancellation or storage refusal can leave a valid smaller result. Showing these states separately helps prevent a large input field from being mistaken for a completed asset.

The public entry permits large targets subject to finite storage and application bounds, but expressly does not claim a completed billion-record generation or simultaneous billion-Gaussian display. More samples of a procedural primitive also do not create newly captured visual information. A denser approximation can improve coverage in some views, yet its evidence remains different from a photographic reconstruction. The detailed storage implementation is outside this report.

Where established research fits

The original 3D Gaussian Splatting paper provides the research context for an efficient Gaussian representation and renderer. Octree-GS studies multiresolution detail selection, while AAA-Gaussians studies filtering, stable projection bounds and culling. These are prior contributions by their respective authors, not inventions attributed to Guassia.

They also show why “more Gaussians” is an incomplete comparison. Representation, view selection and projected coverage affect work and appearance. Guassia’s documented count separation is a product-engineering case study. It does not establish a new level-of-detail algorithm, implementation of either cited method or superiority over research renderers. Preserving source data and reporting effective preview limits can be valuable without asserting any of those claims.

A useful next evaluation

A stronger study would freeze rights-cleared scenes, source files, software versions, camera paths and preview settings. It would vary source count and instance count independently, then compare an all-visible scene with one containing substantial off-camera content. Device, browser, driver, viewport and negotiated limits should accompany every result. Cold load, warm load, memory use, frame-time distribution and image difference should be reported separately.

Large-collection validation would add completed-count verification, cancellation, reopening and storage-pressure cases. Those are proposed checks, not completed findings. The useful publication claim today is narrower: Guassia’s public records describe finite authoring and preview budgets and show why counters need explicit meanings. A trustworthy creator tool lets someone see that distinction before they interpret a large number as a promise.

References

1. Guassia: In-scene creation, larger Gaussian projects and desktop stores — September 23, 2026
2. Guassia: Region editing, source exports and camera-aware Gaussian rendering — September 24, 2026
3. Guassia: A separate Asset Engine, native editing and creator update requests — September 27, 2026
4. Kerbl et al.: 3D Gaussian Splatting for Real-Time Radiance Field Rendering, 2023
5. Ren et al.: Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians, 2024
6. Steiner et al.: AAA-Gaussians: Anti-Aliased and Artifact-Free 3D Gaussian Rendering, 2025