Recently, the research team at Apple Company collaborated with the University of Hong Kong published a new research result titled "Less Gaussians, More Textures: 4K Feedforward Texture Splatting" and introduced a new technical framework called LGTM. This technology aims to solve the computational bottlenecks of current 3D Gaussian Splatting at high resolutions, opening up new paths for graphics rendering in high-pixel display devices such as Apple Vision Pro.

Technical Breakthrough: Decoupling Geometry and Resolution

Although mainstream feedforward 3D Gaussian Splatting technology can quickly convert 2D images into 3D scenes, the computational demand increases exponentially with the rendering resolution, making real-time rendering at 4K level difficult.

LGTM Framework’s core innovation lies in decoupling geometric complexity from rendering resolution. Instead of blindly increasing the number of Gaussian functions, it improves image quality through the following two-step approach:

  1. Geometry Learning: The model first learns the basic structure of the scene from low-resolution images and compares it with high-resolution real images (GT), ensuring that the geometric skeleton is seamless or free of defects at 2K/4K resolutions.

  2. Texture Overlay: A specialized appearance network is introduced to convert fine details from high-resolution images into textures, which are then overlaid on simple geometric shapes.

QQ20260403-100737.jpg

Empowering Apple Vision Pro

The dual display of Apple Vision Pro has about 23 million pixels, with a resolution per eye exceeding 4K. Traditional SPLAT feedforward models often face performance limitations when processing such high-density pixels. LGTM can upgrade existing systems (such as NoPoSplat, DepthSplat, etc.), allowing devices to generate more detailed textures, sharper text, and more realistic immersive environments while maintaining low computational load.

Through project demonstration page, it can be observed that LGTM significantly enhances the visual limits of feedforward 3D reconstruction. This indicates that future visionOS users may experience unprecedented realism in real-time generated virtual spaces.