A generated frame held the pleats on a Doré half-disc through a whole camera move, then dropped one on the way out. We caught it in finish and rebuilt the edge by hand. That is the real size of the problem in front of us.
Artificial general intelligence makes for a good argument over a long dinner. It is a poor basis for a production schedule. The conversation has drifted so far ahead of the tools that it now describes machines nobody in this studio has ever operated.
We run a generative engine most days of the week, for fragrance and jewellery houses that will not accept a soft frame. What we know about these systems, we know from fixing them.
Clients ask, in different words, whether the machines are about to make a studio unnecessary. Almost nobody asks what the tools failed at last Tuesday. Wrong pair of questions, wrong order.
What today's systems are genuinely good at
Range, first. A model will build the world behind a bottle in the time it takes a location scout to answer the phone. That is not a small thing when a house wants desert, then marble, then nothing at all.
Speed second, and speed changes the kind of argument you can have with a client. You stop describing a direction and start looking at it. Someone who could not picture a pleated hoop on raked sand can now see it and say no on sight. A no that arrives in the first week is worth more than a yes in the fourth.
Consistency across a set, third, under supervision. For RIIFFS the bottles are photographed in studio and the world behind them is generated, then composited. NUSUK leans desert and heritage. RIIEFF runs dark-luxe. The engine builds the world; creative direction decides which of the two it is.
Range without a point of view is only volume. The engine produces options no single shoot could reach in a week. It does not know which of them belongs to the house, and that part has not moved at all.
Where they fail, reliably
They do not know your product. A model has no idea what the fluting on a Doré half-disc does under a hard source, because that form exists in the jewellery and nowhere else. Reference, direction and a finishing pass carry it. Nothing else does.
They lose things between frames. An object that has to survive a cut — a patterned cap, a monogram, a pavé setting — drifts, and the drift is small enough to pass a quick look and fail a client's.
They have no restraint. Given room, a model adds: more smoke, more gold, more light. Christian Provenzano's house is built the other way, on a patterned cap, a journal serif, a packshot kept slightly withheld. Every one of those is a subtraction, and subtraction is the thing you have to impose from outside.
None of that is a forecast about next year. It is a description of this week, written by someone who spent it in the finish.
What the AGI conversation actually costs
AI hype is expensive in a way that never shows up as a line item. It damages work in two directions, and both land on the same invoice.
One brand waits. They have read that the systems will handle all of it soon, so they hold a launch film until the tools mature. The launch does not wait for the tools. The season closes, and the film that would have carried it gets made in a hurry by someone else.
Another brand decides the tools have already arrived, buys the cheapest generation on offer, and ships frames with dead metal and a bent serif. The audience never names the fault. They simply don't look twice, and a house that spent years earning a second look has bought itself a first-look brand.
Neither brand was careless. Both were responding to the same noise about artificial general intelligence, in opposite directions.
The useful question was never what these systems will become. It is what this one can hold, in this frame, on this brief, before a person takes it the rest of the way. We answer that per shot, and the answer moves month to month.