Apple releases open-source model that instantly turns 2D photos into 3D views
Model capabilities & applications
- Converts a single 2D image into a 3D-like “spatial” scene using Gaussian splats; demos are widely seen as impressive, especially for photos of people and rooms.
- Expected uses: Apple Vision Pro content, iOS “spatial scenes,” lock-screen effects, real estate imaging, VR experiences, and personal memory enhancement (e.g., deceased relatives, historical footage).
- Some foresee major impact on creative workflows (film, graphics, VR worldbuilding) by dramatically lowering content-creation time.
Technical characteristics & limitations
- Uses Gaussian splats rather than meshes; some users have hacked in mesh export via other projects, but note artifacts and holes when moving the camera off-axis.
- Limited viewpoint freedom: good for small viewpoint shifts / stereoscopy, not full 6DOF; glitches appear with larger movements.
- Resolution and layers are capped, and it doesn’t handle multi-image fusion or robust inpainting of unseen regions.
- Others mention alternative or related projects (StereoCrafter, GeometryCrafter, HunyuanWorld, Marble) that may be better for level design or video/temporal consistency.
Relation to Apple products
- Several commenters believe this (or a close variant) underpins iOS / visionOS “Spatial Scenes” and Photos 3D effects.
- Users of those features report strong emotional impact but also note they are more constrained than the research demos, likely to preserve illusion quality.
Tooling and accessibility
- Some frustration over Conda-based setup; others report the repo’s instructions work and suggest alternatives like
uvor pixi, or just plain virtualenv. - A Hugging Face space makes it usable in the browser; one commenter says Apple isn’t “serious” without an official frontend, others counter it’s a research release.
Licensing and “open source” debate
- Weights are licensed for “research purposes” only; many argue this is not open source under OSI’s “no field-of-endeavor restrictions.”
- Distinction is drawn between open source, source-available, and “open weights”; some say the HN title is misleading, as the repo itself doesn’t claim to be open source.
- Broader frustration that big tech (Apple, Meta, others) market such releases as “open” while retaining commercial control, and that AI threads constantly devolve into definitions of “open source.”
Copyright status and ethics of weights
- Heated discussion on whether neural network weights are copyrightable “tables of numbers,” with conflicting claims and references to database protections.
- Some argue the license mainly signals litigation risk rather than clear legal boundaries.
- Moral questions raised about enforcing proprietary rights on models trained on unlicensed copyrighted data; replies note that if training is legal (fair use), that doesn’t automatically justify ignoring the model license.
Researchers’ origins & US STEM pipeline
- Side thread on many authors having non-US educational backgrounds.
- Points raised: you can’t infer birthplace from names; the US is a small share of world population; foreign-born researchers form a large fraction of CS PhDs and often stay.
- One long analysis frames the situation as universities backfilling funding with international students, creating a pipeline industry later tapped for AI research; scaling a purely domestic PhD pipeline to similar levels is seen as unrealistic.
Broader attitudes toward Apple
- Mixed views: some praise Apple for impactful, user-facing applications of AI; others distrust its closed ecosystem and see this release as another example of pseudo-openness.
- Historical grievances aired (OpenCL, lack of Linux on Apple Silicon, App Store control), contrasted with Apple’s substantial genuine open-source contributions in other areas.