Inverse Construction Is Geometry

"Anyone can draw a box at a known angle. The hard part is looking at a photo and recovering the angle."

Forward Is Easy, Inverse Is Hard

Every construction problem has two directions. The forward direction takes parameters and produces a picture: give a 3-D renderer a head model, a yaw of 20 degrees, a focal length, and it draws the projected result. Any graphics engine does this, and it is not hard.

The inverse direction is the one that matters for study, and it runs the other way: you have the picture — a photograph of a real face — and you want the parameters back. What yaw, pitch, roll, and camera produced this exact image? Recover those and the whole construction falls out of them. That inverse is the entire job of Loomis, and inverse problems are famously harder than the forward ones they undo.

Why Not Just Ask a Model to Draw It?

The tempting shortcut in 2026 is: train a model on thousands of photos-with-construction and let it draw the ball and box directly. It will produce something that looks like construction. But looking right is not the bar. A generative model optimizes for plausible marks, not for geometric consistency — the ball it paints need not actually agree with the jaw angle it also paints. It imitates the surface of the answer without solving the structure underneath. For a tool whose entire value is being the trustworthy reference, imitation is disqualifying.

An Answer Key Must Be Reproducible

Here is the sentence the whole engine is built to protect: an answer key you can't reproduce isn't an answer key. Feed geometry the same photo twice and it returns the identical construction, to the decimal, forever. It is a pure function of the input. A model with sampling temperature, weight updates, or any drift cannot promise that — and the moment your reference answer changes between Tuesday and Thursday, it stops being something you can grade yourself against. Geometry is chosen here not because it is fancier, but because it is faithful.

Geometry Can Say "I Don't Know"

There is one more quiet advantage. When a face is turned too far, or the lighting hides the landmarks, the geometry doesn't bluff — it produces a high reprojection error, a number that says this fit is shaky, trust it less. A generator, by construction, always outputs something confident-looking even when it has no idea. An honest "I'm not sure" is worth more to a student than a confident wrong answer, and only the geometric path can give it.

Code

The two directions, side by side·python

# FORWARD: parameters -> picture. Easy. Any 3D renderer does this.
def render(head_model, yaw, pitch, roll, focal):
    posed = rotate(head_model, yaw, pitch, roll)
    return project(posed, focal)          # a picture

# INVERSE: picture -> parameters. Hard. This IS Loomis.
def recover(photo):
    pts2d = detect_landmarks(photo)       # where the features actually landed
    pts3d = CANONICAL_HEAD                 # where they sit on a reference skull
    yaw, pitch, roll, focal = solve_pnp(pts2d, pts3d)   # invert the projection
    err = reprojection_error(pts2d, pts3d, yaw, pitch, roll, focal)
    return (yaw, pitch, roll, focal), err  # the answer AND how much to trust it

# A generator would skip straight to drawing marks that look plausible.
# recover() instead SOLVES for the parameters -- reproducibly, with an error bar.

Exercise

Name a forward/inverse pair from your own field — render/parse, encode/decode, compile/decompile, synthesize/analyze. For your pair, which direction is harder, and why is the answer to the hard direction sometimes non-unique? Then argue whether you'd want that hard direction solved by an exact method or approximated by a learned model, given you need to trust the result.

Hint

Forward problems evaluate a function; inverse problems invert it, and inverses are often ill-posed — many causes can produce the same observation (many head angles can look similar in one photo, which is exactly why Loomis has to bound its focal search). 'Trustworthy and reproducible' usually pushes you toward the exact method, even when a model is easier to build.