A 3D scene you can repaint, reshape, and remix like a real object

AI can now turn a handful of photos into a 3D scene you can fly a virtual camera around. But those scenes are baked solid. You can't easily reach in and recolor one chair, reshape a vase, or swap a wood texture for marble. Editing them has mostly meant crude moves: shifting a whole object rigidly, or tricks that only work for one type of object.

NeuMesh++ fixes this by storing the scene's look on the corners of a normal 3D mesh — the same triangle skeleton used in video games. At each corner it keeps three separate notes: one for shape, one for color and material, one for meaning ("this is the cup, that is the table"). Keeping these three notes apart is the key. You can rewrite the color note without disturbing the shape, so painting a surface doesn't warp it.

That split unlocks real editing. Drag the mesh to bend the geometry. Paint, fill, or swap a texture onto a chosen patch — a learnable "correction color" at each corner keeps the new texture sharp instead of muddy. And because every region carries a meaning tag, you can select "the cushion" by name instead of hand-tracing it.

The paper shows examples on real and synthetic scenes, but offers no head-to-head accuracy numbers here, and editing still leans on a decent input mesh.

Why you should care: This is the difference between a 3D scan you can only look at and one you can actually edit — recolor one object, reshape another, without rebuilding the whole scene.