ExplorXR

XR Daily: Apple Skips Spatial Capture

· 7 min read

Apple’s $1,999 iPhone Duo Can’t Shoot Spatial Photos or Video

Apple’s most expensive new phone has two rear cameras and still cannot capture spatial content. The iPhone Duo lists a 48MP Fusion Main and a 48MP Fusion Ultra Wide, but spatial photos and spatial video are missing from its supported camera features. The iPhone 18 Pro models list both, with spatial video at 1080p 30fps, and Apple has not explained the omission.

It is not a processing limit. The Duo runs the same A20 Pro chip as the 18 Pro pair, including the dual 16-core Neural Engine, and Apple says the Ultra Wide module is identical. Spatial video has been a first-class feature since the iPhone 15 Pro in 2023, which makes dropping it from a flagship odd.

The consequence lands on the XR side. Vision Pro converts ordinary 2D photos into spatial photos with machine learning after the fact, Quest 3 turns browser images into 3D, and Instagram on Quest uses AI view synthesis for video. That estimates depth and synthesizes a second view, a different pipeline from recording through two lenses. If the priciest iPhone stops feeding the native pipeline, more headset spatial content gets generated instead of captured.

UploadVR


WebAssembly Runtimes in 2026: A Boring Benchmark That Matters

Frank Denis published the 2026 edition of his libsodium WebAssembly benchmarks, which ask one unfashionable question: compile the same C code to wasm, run it on runtimes from this year, last year, and two years ago, is anything actually faster? The numbers are geometric means of slowdown against a native x86-64 build, measured on an AMD Ryzen AI 9 HX 470 with boost disabled at 2GHz.

The tiering is the story. WAMR in AOT mode sits at 1.57x native, unchanged from 1.59x in 2025. WasmEdge reaches 1.74x once you remember the --run-mode=aot flag, because its default path silently interpreted the modules. Wasmtime improved every year, from 2.67x to 2.54x to 2.41x. Wazero is flat near 4.7x, Node moved from 8.60x to 7.95x, and Bun is three times faster than in 2025 but still behind Node. Wasmer recovered in 2026 without beating its own 2024 baseline.

For anyone shipping WebXR or browser graphics tools, architecture dominates version bumps. Ahead-of-time compilation buys a 3x to 4x gap over JIT paths, and SIMD plus the new wide_arithmetic instructions matter for rendering math as much as for crypto.

Frank Denis


RayNeo’s Camera-Free Glasses Land, as Oslo Bans Smart Glasses in Schools

RayNeo, the TCL subsidiary, launched the iO Smart Glasses with a display, no camera, and no speakers. That leaves a 33-gram frame that nearly reads as normal eyewear, with a dual monochrome waveguide at just over 23 degrees of field of view and 1200 nits peak brightness, readable even against a sunny sky in SkarredGhost’s testing.

The tradeoffs are concrete. The display sits high in your vision, fine for notifications but straining for long reads like the teleprompter, and it makes you look like you are staring upward mid-conversation. No privacy coating means bystanders can read your lens, and no speakers means no music and no calls, so the headline Ray-Ban Meta use cases are gone. What remains: a 4-mic array plus a bone conduction mic that tells whether audio came from you or someone in front, and battery life of 40% drain across 8 hours of heavy use.

The social context arrived the same week. Oslo banned smart glasses in schools for the whole school day at every level, effective August 27th. Vice Mayor Julie Remen Midtgarden said the city refuses to let children become “extras in Meta’s data collection,” and noted Meta has reportedly built facial recognition for its glasses. A display without a camera may be the only configuration that survives that scrutiny.

The Ghost Howls · The Local


Impermanence: Rooms That Rearrange When You Look Away

A one-week game jam entry called Impermanence does something cheap to build and deeply unsettling in a headset: rooms change every time you stop observing them, and blinking counts too. Developer liamflannery56 built it in Godot for the Brackeys jam, which ran August 23rd to 30th under the theme “Trust No One.”

The technique is simple. The environment does not simulate hidden state; it marks whether a door has been seen before, and when the camera raycast hits the edge of the doorframe, the room behind it swaps. An outline shader supplies the eerie look. That is a stencil-style portal trigger dressed as an observation mechanic, and it works because the player’s attention is the clock.

It is a pattern worth stealing for VR horror and puzzles, where gaze and head pose are already tracked every frame. Free on itch.io.

80 Level


DeepSeek’s V4.1 Flash Undercuts Its Own Pro Model

DeepSeek shipped V4.1 Flash around September 10th claiming it beats V4 Pro on performance, cost, speed, and task completion time, then went further than most labs would: every Pro request is being routed to V4.1 Flash and billed at Flash pricing until V4.1 Pro lands. Rates are $0.003 per million input tokens on cache hits, $0.15 on cache misses, and $0.60 per million output tokens off-peak, doubling at peak.

The Hacker News thread is the more useful artifact. Some developers report language drift plus infinite loops and invalid tool calls inside coding harnesses. Others argue the pathology comes from third-party providers running lower quants while the official API stays fast. The most practical note for anyone automating a pipeline: replay anonymized sessions against your harness and fix it until the only failure left is the model.

Token math matters for XR. Assistant features in headsets are billed per interaction, and cheaper inference is what makes it reasonable to keep an agent awake rather than waking it on demand.

Hacker News · DeepSeek Pricing


A MetaHuman, ZBrush, and Marvelous Designer Pipeline for Dr. Doom

Patrick van Rooijen, a senior character artist at Rocksteady with Horizon Forbidden West, Burning Shores, and Suicide Squad: Kill the Justice League behind him, broke down his Dr. Doom project for 80 Level. The head started as a scan of his cousin, converted to a MetaHuman mesh in Unreal Engine, taken into ZBrush for a stylization pass that reshaped the silhouette and added scarring, then brought back into Unreal to update the MetaHuman shapes.

The costume follows a familiar real-time workflow: a ZBrush blockout to establish layering and construction, then a rebuild in Marvelous Designer, with the neck collar and layered cape the hardest fit. His rule is worth repeating: understanding how a garment is built and where tension comes from produces believable folds, while sculpting decorative wrinkles does not. Hard-surface pieces were cleaned with ZRemesher and HPolish, retopologized in Maya and TopoGun, UV’d in RizomUV, and textured in Substance 3D Painter.

80 Level


Quick Bites

  • A free Blender tool auto-generates deform bone chains along stylized hair strands, with control over bone count, root depth, and direction. Needs Blender 4.5 LTS or newer. → 80 Level
  • A Godot 4 port of Bethesda’s 1996 XnGine shooters, SkyNET and Future Shock, rewrites the engine while reading BSA archives, WLD heightmaps, and HMI music from your own copy of the original data. → GitHub
  • Meta Connect 2026 runs September 22nd to 24th, and Phoenix prescription lens inserts look like a Zenni partnership. → Road to VR

Related reading