At today’s WWDC 2025 keynote, Apple dropped two announcements that might hint at something big under the hood, radiance field-style representations powering spatial media. While the company didn’t use the term directly, their new Spatial Scenes and upgraded Personas in visionOS 2.6 sound suspiciously like they’re borrowing from the same tech stack behind NeRFs and Gaussian Splatting.
Apple introduced Spatial Scenes, a new way to bring photos to life using “a new generative AI algorithm and computational depth.” These scenes give users multiple perspectives on a photo, making them feel like they can lean in and look around, a hallmark of radiance field-based reconstruction. Combined with multi view depth and ML, this sounds more than just parallax. It’s interactive 3D from a flat source, likely built on a learned scene representation. Zillow is already using the API in their Immersive app, giving apartment photos realistic depth and volume.
The upgraded Personas feature showcases “industry-leading volumetric rendering and machine learning,” creating expressive avatars with full side profiles, realistic hair, and high-fidelity skin tones. Generated on-device in seconds. This level of photo realistic, view consistent rendering at runtime fits squarely into the realm of radiance fields, especially considering Apple’s emphasis on both realism and lightfield like rendering flexibility.
Apple has never been one to name their tech stack, but the language here is striking. "Volumetric rendering," "generative AI depth," and "viewing scenes spatially" are all phrases lifted straight from the playbook of radiance field research. Whether it’s radiance field inspired or some optimized proprietary version, it’s likely that Apple is using neural rendering under the hood and these spatial features might be its most public expression yet.
We’ll be watching closely as developers begin to poke around the new Spatial Scene API. If it exports or supports radiance field like data, we might have just witnessed Apple’s first mainstream deployment of learned scene representations.
