Existing 3D Gaussian scattering methods predict pixel-aligned primitives, resulting in a quadratic increase in the primitive count as the resolution increases. This fundamentally reduces their durability, making high resolution integration such as 4K impossible. We present LGTM (Less Gaussians, Texture More), a feed frame that overcomes this resolution scaling barrier. By predicting congruent Gaussian primitives combined with per-primitive textures, LGTM separates the geometric complexity from rendering resolution. This method enables the synthesis of a novel high-fidelity 4K scene with no per-scene optimization, a capability previously unattainable in forward methods, all while using very few Gaussian primitives.
- † University of Hong Kong
- ** Work done while at Apple
