How to Rebuild a Flower Out of Light

We have written before about what the macro lens knows, which is everything, and what the hero shot admits, which is nothing. The argument was simple. A photograph you cannot turn is an alibi, and a 360 spin is a cross-examination, and good flower has nothing to fear from either. That piece ended with the bud spinning. This one is about what happens when the spinning stops, which is, if you do it properly, the strangest and most beautiful thing in product imaging right now. You take the flower apart into light and you put it back together inside the computer, whole, solid, and turnable forever.

The trade calls it gaussian splatting, a term with all the romance of a dropped pie, so let us ignore the name and describe the meal.

It starts the way everything here starts, with a real flower and an unreasonable number of photographs. The bud sits on a turntable in a black void, lit like relic bone, and the macro rig fires as the table creeps through one slow revolution. By the end of it we have seven hundred frames of a single nug, each one six thousand pixels wide, which is more portraiture than most heads of state receive in a lifetime, lavished on an object that will eventually be set on fire in someone’s garage.

Then the flower goes home in its jar and the photographs go to work.

The first program does something that sounds impossible and is merely tedious. It reads the photographs, only the photographs, and works out precisely where the camera was standing when each one was taken. Every frame, placed in space, accurate to a fraction of a pixel. This is not magic. It is triangulation, the same dull, immortal math that surveyors and artillery officers have trusted for centuries, done a few million times without complaint. When it finishes, we have something no photographer has ever actually owned before: a perfect map of the shoot itself, every camera position pinned like a butterfly.

Now the reconstruction. A second program fills the empty dark with millions of tiny translucent daubs of colour, little airborne brushstrokes, and then begins the least glamorous job in art history. It compares its daubs against the photographs, from one angle, then another, then all seven hundred, and every daub that disagrees with the photographic record gets moved, shrunk, recoloured, or executed. It does this over and over, tens of thousands of rounds, until the painted thing and the photographed thing cannot be told apart from any direction we ever pointed a lens.

Notice what is missing from that description. Nothing is imagined. Nothing is decorated. There is no step where the flower gets flattered. Every speck in the finished model holds its job because it matched a photograph of the actual bud, and any speck that could not is gone. It is the least imaginative painter who ever lived, and that is exactly why you can trust it. Your flower does not come out improved. It comes out true, down to the individual trichome heads, standing up like glass pins.

Here is the part that would have made an old advertising man weep into his third martini. The computer that does this heavy lifting is not ours. We rent it, by the hour, from a warehouse somewhere in the great humming middle of nowhere, and the entire reconstruction of one flower costs about what a gas station lighter costs. Twenty-odd cents an hour for the kind of machine that would have filled a university basement when I was learning to read. The whole apparatus of the miracle, rented like a carpet cleaner. This is either the triumph of civilisation or its punchline, and on this question we remain agnostic.

A confession about the leftovers. The process saves its homework as it goes, and the early saves are gorgeous failures: the bud half-formed, dissolving at the edges into green weather, a flower photographed mid-seance. We used to throw these away. We do not anymore. Run the sequence from fog to solid and you get the flower assembling itself out of nothing, which cuts to a beat drop better than almost anything we own. The mistakes turned out to be the special effect.

And then the finish, which is where the reconstruction stops being a laboratory exercise and starts being an image. Because once the flower exists inside the machine, the camera is freed from every law that governed it on the shoot. We can orbit the bud slowly, like something in state. We can push the lens into the gap between two calyxes, a space no physical camera could enter without a crime being committed against the product. We can light it, float it, dissolve it, run colour through it like weather over a mountain range, and render the whole performance at 4K before it is finished by hand, frame by frame, in the edit. The flower becomes an actor with infinite patience and no rider.

The honest question a dispensary owner should ask is the same one they should ask about everything we do: what is this for? And the answer is the same answer as always, with the volume turned up. A hero shot can lie. A 360 cannot stop turning, so it lies less. A full reconstruction cannot lie at all, because it is assembled from every angle simultaneously, and a bad side would be built into the model with the same devotion as a good one. Which means the flower has to be worth it before the first frame fires. This pipeline is a privilege of clean product. It is the moat again, deeper. The brand that can hand a stranger its flower as a solid, turnable, enterable object is saying the loudest thing a brand can say, which is: look anywhere.

The shelf is flat. The jar is smudged. The screen, it turns out, no longer has to be either. We photograph the flower seven hundred times, we let the mathematics argue about it for half an hour, and what comes back is the thing itself, portable, permanent, and unable to keep a secret. Good flower deserves nothing less. Bad flower, as ever, should be very worried indeed.