Splat Report weekly, 2026-W28

PlayCanvas engine v2.20.0, released 23 June, replaced the compute-based gsplat renderer with a new hybrid approach and added shader customisation hooks including gsplatModifyPS to complement the existing gsplatModifyVS. The six point releases that followed, v2.20.1 through v2.20.6, landing between 29 June and 6 July, spent that fortnight fixing the fallout: a crash when unloading a SOG asset still referenced by the streaming world, a SOG bundle parser that kept running after an asset unloaded mid load, SOG streaming reloads returning destroyed textures from the loader cache, a GPU device hang when picking gsplats on Windows and NVIDIA hardware, and several gsplat component teardown crashes.

None of this changes the SOG format itself or how a scene is authored. It is a straightforward reminder for anyone pinning the engine version, as this pipeline's own documentation recommends: v2.20.0 was not the release to sit on for production SOG delivery, and v2.20.6 is where the hybrid renderer's rough edges settle.

Anthropic launched Claude Sonnet 5 on 30 June, with a one million token context window that is not an optional extended mode but the model's only context size, alongside a 128,000 token maximum output. Anthropic positions it as closing most of the capability gap with the pricier Opus 4.8 on coding, reasoning and tool use, at introductory pricing of $2 per million input tokens and $10 per million output tokens through 31 August, rising to $3 and $15 after that. It is now the default model on the Free and Pro plans.

For anyone running long agentic coding sessions, a genuinely large native context changes how much project history, prior tool output and accumulated conversation an agent can hold without summarising or losing earlier decisions. The same week's Claude Code changelog also moved background subagents to run by default, which compounds with the larger context window for longer, less supervised agent runs.

Arrival.Space's 2026.6 release adds Google Photorealistic 3D Tiles as a streamable base layer, plus a Splat/Tile Cropper for trimming a capture to fit the surrounding tiled geometry, and opt-in WebGPU rendering. The practical effect is compositing a Gaussian splat capture into a live, streamed geographic context rather than an empty scene, useful for the platform's social sharing use case of placing a scan somewhere recognisable on a map.

This isn't part of this pipeline's own delivery stack, which stays inside PlayCanvas rather than a geospatial tile viewer, but it's a useful data point on where browser-based splat viewers are heading generally: WebGPU adoption and tiled, streamed geographic context are both directions the wider ecosystem is converging on independently of any one vendor.

Screen-space artefacts, sensor defects, rain or mud on a lens enclosure, a thumb over the sensor, a watermark, are fixed to the 2D image plane rather than the 3D world, but standard 3D Gaussian Splatting has no way to know that and bakes them into the scene as floaters or near-camera geometry. SSA-3DGS jointly optimises the 3D scene alongside a learnable 2D overlay, exploiting the fact that a real artefact stays static in image space across every view while the true geometry changes correctly with viewpoint. That geometric disagreement is enough to separate the two without supervision or manual masking.

Across synthetic corruptions and a self-captured real-world dataset, the authors report up to 9dB PSNR improvement over ordinary 3DGS trained on the same corrupted inputs. No code release was mentioned in the paper, so it's a research method rather than a tool available today, but the failure mode it targets, a stray reflection or dust speck baked in as a permanent floater, is a familiar one from real captures.

YaGS is an open-source Unreal Engine 5 plugin from Yandex, supporting UE 5.5 through 5.7, that brings a Gaussian splat capture into a level as a proper scene object rather than a standalone viewer asset. The specific problem it addresses is depth: a splat has no surface topology or collision geometry to test against, so naively compositing one with ordinary UE geometry tends to produce z-fighting or objects sinking into or floating above the capture. YaGS handles that depth interaction correctly, so a scanned environment and hand-modelled geometry can occlude each other the way two solid objects would.

Yandex archived the repository on 2 July, the same day it announced it, so it ships as a reference implementation of correct splat/mesh depth handling rather than a maintained plugin. For archviz practice, it is still worth studying the approach: this is the CG-compositing problem this pipeline's own PlayCanvas delivery doesn't yet have to solve at the engine level, since splats and meshes there are usually kept as separate concerns rather than mutually occluding in one scene.

OctaneRender 2026.4 adds support for SPZ v4, the compressed Gaussian splat format Niantic released in May, alongside reduced GPU memory use for splat scenes, fixes for rendering scenes above 30 million Gaussians, and Neural Radiance Cache support on macOS. It landed the same week as Foundry's Nuke 17.1 open beta, which adds animated splats and basic relighting, another sign that mainstream rendering and compositing tools are treating Gaussian splats as an ordinary asset type rather than a novelty import.

None of this touches this pipeline's own SOG delivery format directly. SOG and SPZ remain separate formats built for different jobs: SOG for PlayCanvas web delivery, SPZ as the format most likely to turn up when a scene needs to leave this pipeline's ecosystem for a renderer or viewer that doesn't speak SOG. Wider SPZ v4 adoption across tools like Octane is exactly the trend that keeps that format worth keeping in mind.

3D Gaussians can be ray traced as well as rasterised, but ray tracing them has so far been far slower, with the existing 3D Gaussian Ray Tracing approach taking nearly an order of magnitude longer to optimise than standard rasterisation. GRay, from Yohan Poirier-Ginter, Jean-Francois Lalonde and George Drettakis, a co-author of the original 3D Gaussian Splatting paper, closes most of that gap by exploiting a property specific to ray tracing: a ray only needs to evaluate the Gaussians it actually intersects, which scales far better than rasterisation as scenes get denser.

The authors report GRay renders nearly 4x faster and optimises nearly 10x faster than prior 3D Gaussian ray tracing, with speed competitive with ordinary rasterised 3DGS at somewhat lower quality. Ray tracing isn't part of any browser splat pipeline today, this pipeline included, but it's the route to correct reflections and multi-bounce light that rasterisation structurally can't do. The paper was published at I3D 2026 in May, ahead of this month's arXiv posting.