A new research framework keeps AI-generated 3D views from a single photo honest about shape, instead of letting them warp into geometric mush.
The method, called GenNVS, tackles single-image novel view synthesis - generating new camera angles of a scene from just one photo. It separates foreground objects from the background and rebuilds each with 3D Gaussian Splatting, a point-based rendering technique, then aligns the two through a coarse-to-fine optimization process into one unified 3D scene. That scene conditions a video diffusion model through a technique the researchers call Dual-Stream Masking, which combines rendered validity masks with geometry-aware warping to guide what gets generated. The researchers report it performs favorably against recent methods on both visual quality and geometric accuracy, while also supporting scene editing.
Novel view synthesis has been stuck between two bad options: multi-view 3D reconstruction, which needs many input photos, and diffusion models, which can produce a plausible-looking new angle while getting the underlying geometry wrong. By building separate 3D representations for foreground and background before handing things to the diffusion model, GenNVS tries to keep generation structurally grounded rather than just pattern-matching pixels - a distinction that matters for anything downstream that needs the geometry to actually be correct, like 3D asset creation or AR placement, not just an image that looks right in a demo.
It is a benchmark result, not a shipped product, and the real test is whether that geometric discipline survives contact with the cluttered, badly lit photos ordinary people actually take.