Skip to content
← Journal

AI Consistency Is the Hard Problem

August 4, 2025 · 4 min read

The bottle has a seam. Once someone in the room notices it, they cannot stop noticing it, and neither can the client.

Somewhere around the fourth frame the cap loses a facet. By the ninth the glass has changed thickness and the engraving sits a millimetre high. Nobody can say when it happened, and the engine cannot be asked, because it holds no memory of the fourth frame while it makes the ninth.

Generation was never the constraint. One beautiful frame of a fragrance bottle is a solved afternoon. Thirty of them, scrolled in sequence, believed by a customer to be the same object, is where the schedule actually goes.

Three places consistency breaks

AI consistency fails in three places: the object, the face, and the world. They fail differently and they want different fixes, and treating them as one problem is how a set ends up half-right.

Object drift is measurable. Put two frames side by side and point at the cap. Face drift is not measurable and is caught faster, because a change too small for anyone to name still registers as a different person. World drift is the slowest and the most expensive: nothing is wrong with any single frame, and the set still doesn't look like one house.

Ranked by how much production time they cost: face, object, world. Ranked by what they cost the brand when they go wrong: world, face, object. The two lists disagree, which is the reason this gets decided before anyone generates anything.

One bottle

The hardest object is the one the customer can hold. A fragrance bottle is a manufactured thing with a fixed silhouette, a known cap, a specific weight of glass, and a client who knows it better than anyone in the room.

Which is why the RIIFFS bottles are photographed in studio and the world behind each one is generated, then composited in. That reads as an aesthetic preference. It is a consistency decision. It settles which half of the frame is allowed to drift. The bottle is fixed because it was shot; the desert behind NUSUK and the dark-luxe behind RIIEFF can be built, because no one holds a reference photograph of them.

Doré is the same argument in jewellery. One fluted, scalloped half-disc returns as a ring, a stud and a hoop, across brushed gold, gunmetal and a pink-violet metal. The signature is a shape, so the shape is what must not move. The metal can change. The pleats cannot. Naming which property is load-bearing takes an hour and saves a week.

One face

Character consistency is the expensive one. The cheapest answer to it is the one Ramón Béjar arrived at for reasons that had nothing to do with any engine: No Faces, Only Storytelling. Architecture sets the frame, motion moves through it, light is built for material, and every scene arrives on surface and colour instead. A brand rule that happens to be a production strategy.

Most houses cannot take that route and shouldn't. When a face is required, the face sets the schedule. It gets locked before the world does — identity first, wardrobe second, environment last — because a world can be rebuilt around a face and a face cannot be rebuilt around a world.

Then it still goes to a person at the end. Hairline, hands, the join at the jaw, the way an ear behaves in a three-quarter turn. The finishing layer is not a rescue operation. It is scheduled work, budgeted like any other pass, because a set carrying one wrong face is not a set with one weak frame in it. It is a set nobody trusts.

One world

Colour is the cheapest consistency device a brand owns and the most underused. Sospiro sorts by it: indigo, gold and oxblood, velvet bodies, gold caps. One palette holds a whole line together as one house even while each fragrance is allowed to arrive as its own act. A palette is a refusal with a pleasant name.

L'Orientale Fragrances runs the harder version of the same test. Two backgrounds, one house. Dark and cinematic, gold and amber on matte glass, shadow doing most of the work. Then the same bottle on white with nothing behind it. The background changes and nothing else does. That constraint is what carries brand visual identity across the jump from campaign film to product grid, which is the jump most brands quietly lose.

Most of what makes a set read as one set is applied after generation rather than during it. Grade, colour, the last detail pass, done by hand. We would like to write something more impressive here. That is where it happens.

The seam is still on the bottle. We decided early, and in writing, where it goes.