"A beginner-friendly intro to how 3D AI sees shapes: voxel grids, in about 40 lines of Python in Google Colab."
FROM ZERO TO FIRST Training a 3D Generator Without a PhD
3D generative AI looks like it needs a PhD. But the first step is surprisingly small: AI doesn't see shapes, it sees numbers.
Let's prove it. In free Google Colab (no install, no GPU), we'll build a simple character from spheres and then turn it into a voxel grid, the kind of array a neural network can learn from.
Step 1: A character made of spheres
import numpy as np import matplotlib.pyplot as plt # (name, center, radius, color) parts = [ ("body", ( 0.0, 0, 0.0), 1.0, "tomato"), ("head", ( 0.0, 0, 1.4), 0.6, "gold"), ("left hand", (-1.2, 0, 0.2), 0.3, "gold"), ("right hand", ( 1.2, 0, 0.2), 0.3, "gold"), ] def sphere(center, r, n=30): u, v = np.mgrid[0:2*np.pi:n*1j, 0:np.pi:n*1j] x = center[0] + r * np.cos(u) * np.sin(v) y = center[1] + r * np.sin(u) * np.sin(v) z = center[2] + r * np.cos(v) return x, y, z fig = plt.figure(figsize=(5, 5)) ax = fig.add_subplot(projection="3d") for name, center, r, color in parts: ax.plot_surface(*sphere(center, r), color=color) ax.set_xlim(-2, 2); ax.set_ylim(-2, 2); ax.set_zlim(-2, 2) ax.set_axis_off() plt.show() You should see a round tomato body with a gold head and two hands. Change a radius or a color and run it again. That's how you start to feel what the numbers control.
Step 2: Turn it into a voxel grid
A voxel is a 3D pixel. We split space into a grid of tiny cubes and mark each one as filled (inside any sphere) or empty.
def voxelize(parts, R=16): axis = np.linspace(-2, 2, R) X, Y, Z = np.meshgrid(axis, axis, axis, indexing="ij") grid = np.zeros((R, R, R), dtype=bool) for _, (cx, cy, cz), r, _ in parts: grid |= (X - cx)**2 + (Y - cy)**2 + (Z - cz)**2 <= r**2 return grid for R in (8, 16, 32): g = voxelize(parts, R) print(f"{R}x{R}x{R} grid -> {g.sum()} filled cells") Your character is now a tensor of True/False values. Higher resolution means a smoother shape and more numbers to store. This is the format models learn from.
Step 3: See it
g = voxelize(parts, 16) ax = plt.figure(figsize=(5, 5)).add_subplot(projection="3d") ax.voxels(g, facecolors="tomato", edgecolor="k", linewidth=0.2) ax.set_axis_off() plt.show() Try R = 8, 16 and 32 and compare. You'll see why resolution matters.
Where this goes next
Once a shape is an array, the usual machine learning toolbox applies:
- Signed distance fields: store the distance to the surface instead of just filled or empty.
- Training: a loop that guesses, measures the error (the loss) and adjusts.
- Diffusion: add noise to data step by step, then train a model to remove it. Start from pure noise and it can generate new shapes.
I cover this whole path, from the first sphere to training a tiny diffusion-based 3D character generator, in a book I wrote for beginners: From Zero to First Character: Training a 3D Generator Without a PhD. Everything runs in Google Colab.
Disclosure: I'm the author. https://www.amazon.in/dp/B0HSKF3DVM
What would you like to see next: signed distance fields, the training loop or diffusion? Tell me in the comments.
