XR Daily: Blender 5.3 Embraces Gaussian Splats

Blender 5.3 Is Getting Native Gaussian Splat Support

After years of leaning on third-party add-ons, Blender is getting 3D Gaussian Splatting as a first-class import format. Blender 5.3 will read splats from PLY and SPZ files and render them through every Blender render engine, so both Cycles and EEVEE handle them. Export is not supported yet, and the release notes list three known limitations: performance is not where it needs to be, rendering is not fully correct because Cycles and EEVEE operate in linear space, and Apply Transform mishandles scale, rotation, and spherical harmonics.

That last point is the one worth reading carefully if you have shipped splat work. A splat fit carries its own coordinate conventions and its rotation lives in the spherical harmonics rather than in a normal object transform, which is exactly why the transform operator breaks on them. The color problem has the same root: capture pipelines deliver data tuned for a specific display response, and a renderer that linearizes everything before tone mapping will disagree with it until someone writes the conversion.

Blender 5.3 is in alpha until September 30 and the final release is expected on November 10, joining native dispersion in the Principled BSDF and the new custom material lighting nodes. The workflow significance is bigger than the feature list suggests. Looking at splats inside a headset is a solved problem, with mature Quest viewers and Steam Frame capture add-ons. Editing, lighting, and compositing them alongside ordinary geometry inside the tool artists already use was the real gap. Background: Gaussian splatting basics.

→ 80 Level

---

NVIDIA Opens Two Tracks for Writing GPU Kernels in Rust

NVIDIA published CUDA Rust, and it deliberately mirrors the two models CUDA itself offers. The SIMT track, cuda-oxide, is a custom rustc codegen backend that routes #[kernel] functions through Rust MIR, the Pliron IR framework, and LLVM down to PTX. It wants Linux, a GPU at compute capability 8.0 or higher, a CUDA 12.x toolkit, clang with libclang headers, and a pinned nightly toolchain. It is early alpha.

The Tile track, cutile-rs, runs on stable Rust 1.89 or newer against CUDA 13.3 with no custom LLVM, is published on crates.io, and is already used in HuggingFace’s Grout inference engine and mistral.rs. Both enforce memory safety at compile time, cuda-oxide through DisjointSlice and launch contracts that prevent aliasing, cutile-rs through tensor partitioning and ownership guarantees.

For graphics and XR work, the Tile model is the more interesting one. It lets you describe what a tile of data does and leaves thread mapping and memory layout to the compiler, which is a much closer fit to how tile-based mobile GPUs actually execute than raw SIMT thread management is. Splat sorting, foveated rendering, passthrough compositing, and tracking inference on a Quest-class chip are all per-frame kernel problems where the scheduling decision is the optimization.

The catch is maturity. cuda-oxide needs a specific nightly and is explicitly alpha, and real production shader work still lives in C++ and HLSL. NVIDIA says inter-language interop is planned so picking Rust does not lock you out of the others.

→ NVIDIA Developer Blog

---

A 93-Gram Tethered Headset Raises $1.2M

Unseen Reality closed its Kickstarter for the URXR One at $1,288,712, having blown past a $50,000 goal within minutes of launch on August 4. The hardware sits in an unusual category: it is not birdbath AR glasses in the Viture or Xreal sense, but a lightweight passthrough headset with no compute and no battery, tethered over USB-C DisplayPort to a laptop, phone, tablet, or gaming handheld. Background: thermal design and compute offload.

That decision buys the spec sheet. The URXR One weighs under 93 grams against roughly 515 grams for a Quest 3, using mini pancake optics and dual 1.03-inch micro-OLED panels at 2,448 x 2,064 per eye, which works out to about 36 pixels per degree at 90Hz. Field of view is 80 degrees horizontal by 56 vertical, or 90 degrees diagonal. Passthrough runs at 2,328 x 1,728 per eye at 90fps with latency under 10ms, driven by four forward cameras and 6DOF SLAM with a 1000Hz IMU. Background: how passthrough works.

The trade is obvious: a cable and a narrower field of view than a standalone Quest. Early backers paid $800 against a $900 list price, with backer shipping targeted for mid-October. This is a monitor replacement first, and $1.2 million from backers says the market for light display glasses keeps growing even as standalone headsets get heavier.

→ Road to VR

---

Vision Pro Has a Rhythm Game That Writes Its Own Levels

Flow Chaser launched on Apple Vision Pro alongside visionOS 27, and it is built around a system capability rather than a hand-authored song list. The game uses Apple’s new on-device music intelligence to analyze the structure of a track and generate a level from it. It ships with 140 original tracks totaling more than 7.5 hours of music, but players can also import DRM-free songs and get a generated level for any of them.

That solves the problem that has always capped VR rhythm game catalogs. Beat Saber-style content is hand-mapped per track, which makes every new song an authoring cost. Procedural generation from on-device audio analysis turns that into an inference cost, and keeping the analysis on the headset avoids shipping your music library to a server.

The lineage is explicit. Masaya Matsuura’s Vib Ribbon let players swap a PlayStation disc for virtually any music CD in the late 1990s and generated levels from whatever was playing, a technique Sony patented. Flow Chaser is that idea running on hardware that can look at the whole audio spectrum in real time, and it is a signal that audio-reactive spatial content is now a system-level API rather than a research demo.

→ UploadVR

---

85.8% of Japanese Game Developers Now Use Generative AI

A preview of the CESA Video Game Industry Report 2026, distributed at Tokyo Game Show, puts generative AI use among Japanese game developers at 85.8% at least occasionally, with 63% saying they use it routinely. Last year’s report had roughly 51%, which is a steep jump for a single year. The full methodology lands in early December.

The number that matters more than 85.8% is what the work actually is. The 2025 report found visual and image generation as the most cited application, followed by story and text generation and programming support, and about 32% of companies reported using AI while developing proprietary engines. Workflow efficiency and shorter schedules were the top expected benefits, and the most common management measure was requiring people to review, correct, or supervise AI-assisted output.

That governance pattern is the useful part for anyone building production pipelines. Tools that produce reviewable, correctable output win; tools that produce a confident result with no intermediate steps do not fit the policy studios are writing. CESA’s managing director Tsutomu Masuda also flagged unresolved copyright and rights issues, which is the same open question hanging over every capture and reconstruction pipeline.

→ 80 Level

---

Night City Was Built by Hand, and the Texel Budget Shows It

CD Projekt Red’s art leadership gave an interview on how Night City was assembled, and the headline claim is that it was done almost entirely by hand, with procedural tools used only as starting points in a limited number of cases. VP and Global Art Director Jakub Knapik and Environment Art Director Kacper Niepokólczycki describe four chronological visual styles across the districts, a rule of contrast used to keep them visually distinct, and the work of creating controlled chaos that never overwhelms the player.

Two production numbers stand out. Cyberpunk 2077 needed twice the texel density of The Witcher 3, and the shift to first-person perspective changed how environments were composed. They also say the studio’s physically grounded lighting pipeline is why path tracing felt like a natural evolution rather than a reinvention, and that clearer visual guidelines made Phantom Liberty more detailed but less noisy.

For XR developers this reads as a warning about viewpoint freedom. Texel density and hand-placed detail are budget decisions made against a camera that the player almost never controls. Take that same scene and let the viewer stand anywhere, and the surfaces that were only ever seen from eye height have to hold up from the floor, from behind, and from inside arm’s reach. It is the same reason splat relighting and dynamic material systems matter so much for volumetric content.

→ 80 Level

---

Quick Bites

• Meta’s Connect keynote may demo Phoenix with a full-body hologram call. The Information reports Meta could show its next-gen headset on September 23 or 24 with a photorealistic Codec Avatar call driven by CTO Andrew Bosworth appearing as the hologram. UploadVR notes the report mislabels Phoenix as AR glasses when it is a passthrough mixed reality headset. → UploadVR
• Spymaster leaves Early Access on October 15. Innerspace’s time-rewinding spy adventure ships three new missions on Quest and Steam, including a finale that puts all three agents on the field at once. → UploadVR
• A 19-year-old environment artist broke down her Ankara alley scene. Nore Pollentier reverse-searched her reference photo to find the exact street on Google Maps, then built a Blender blockout, a modular kit, and Substance 3D Designer tiling materials for Unreal, which is the modular workflow that keeps XR scene budgets manageable. → 80 Level
• flat.social shipped a new build of its browser spatial meeting app. Avatars walk between rooms with distance-based spatial audio and no install, and the release pulled 97 points on Hacker News. → flat.social