Generative 3D AI & single-image reconstruction

A generative model's first guess is rarely good. Here every point starts collapsed onto the same spot — the shape's centroid — and gets refined step by step. Step through it: the point cloud fans back out toward the 4 corners it's supposed to represent, and the fidelity bar climbs as it goes.

"Good enough" needs a metric. Chamfer distance scores how well a predicted point set $A$ matches a target $B$, symmetrically — it penalizes both a predicted point sitting far from any real surface point, and a real surface point that nothing predicted is near:

Once a shape is refined enough, turning it (or a continuous density field) into an actual mesh means finding where the field crosses zero — the surface. On a 1D slice, that's linear interpolation between two samples of opposite sign:

• $A, B$ — the predicted point set and the target (or vice versa); the two terms make the metric symmetric.
• $d_0, d_1$ — a sampled field's value at two adjacent points $x_0, x_1$ (negative = outside, positive = inside).
• $x_{\text{surface}}$ — where the field crosses zero between those two samples — marching cubes' core 1D building block.

1. Coarse guess: all 4 points collapsed onto the centroid (1,1)

   Every point is $\sqrt{2}$ from every one of the target square's 4 corners in both directions: chamfer $= \sqrt{2} + \sqrt{2} = 2\sqrt{2} \approx 2.828$.

2. Halfway refined

   Each point has moved halfway to its corner: chamfer drops to exactly $\sqrt{2} \approx 1.414$ — half the coarse score.

3. Fully refined: points exactly on the corners

   chamfer $= 0$ — every predicted point sits exactly on a real surface point, and vice versa.

Drag $t$ continuously instead of jumping between 3 fixed steps. Chamfer distance falls off perfectly linearly here — $2(1-t)\sqrt{2}$ — because every point is moving in a straight line toward its own target corner at the same rate.

Locating a surface crossing in a 1D density profile sampled at $x = 0, 1, 2, 3$ with values $-1, -0.2, 0.6, 1.5$ (negative = outside, positive = inside):

1. Between x=1 (d=-0.2) and x=2 (d=0.6): opposite signs, a real crossing

   $x_{\text{surface}} = 1 + \frac{0.2}{0.6-(-0.2)}(2-1) = 1 + 0.25 = 1.25$ — the surface passes a quarter of the way from $x=1$ to $x=2$.

2. Between x=0 (d=-1) and x=1 (d=-0.2): same sign, no real crossing

   The formula still returns a number ($1.25$ again, by coincidence of these particular values), but it falls outside the segment $[0, 1]$ — the tell that there's nothing to mesh here, since both samples agree they're outside the shape.