Bihe Xi has released PinchDrop v0.5.1, a Vision Pro app that unfolds a gaussian splat scene from a photo on a phone when the wearer pinches and opens a hand.
Each PinchDrop scene pairs one reference image with one prepared Gaussian model. When a scene is selected, the headset registers that image and its physical display size with ARKit, tracks the photo and the hand, samples the photo's colors and renders the paired scene locally. Nothing is trained, recognized in the cloud or transferred during the pinch.
The Xcode project ships four scenes with their phone photos, a Beijing ginkgo, Seljalandsfoss and two snowy Irkutsk streets, the scenes from the demo videos Xi has posted on X, including one on November 19. A PinchDropPreview target previews the reveal on an Apple silicon Mac, and the hand interaction needs a Vision Pro on visionOS 26 with Developer Mode. Rendering runs on a vendored copy of Sean Cier's MetalSplatter, the MIT licensed Swift renderer that already runs splats on Vision Pro.
Scene generation stays outside the app. A bundled script pairs a photo with a SOG v2, SPLAT, SPZ or compatible PLY scene and caps each scene at 650,000 Gaussians by default. The README covers making scenes with a free Mac app or, for noncommercial research, Apple's ML-SHARP model and shrinking them with PlayCanvas's splat-transform.
PinchDrop follows Walk Into Memory, the MIT licensed photo to splat pipeline and visionOS app Xi open sourced on October 28, which also builds on MetalSplatter and adds hand gestures that dissolve splats near the hand and set the scene swaying in a wind.
PinchDrop's original code and documentation are source-available under the PolyForm Noncommercial 1.0.0 license, and the four photographs and paired scenes are supplied for noncommercial use with the app. Apple licenses the ML-SHARP weights for noncommercial scientific research only, excluding product development, and PinchDrop does not distribute them.
PinchDrop is available now on GitHub.




