The image problem was reaching back into the recipe.
Culinary sourced roughly 1,450 images a year — approximately, as an internal operating estimate — across DAM sourcing, stock libraries, studio photography, and manual Photoshop work. Stock read as generic or plainly wrong for the dish. Studio shoots were slow and expensive.
The part that mattered was further upstream. When the right imagery was not available, the recipe was sometimes changed to match the photography that was. That is not an efficiency problem. That is a production constraint editing the product.
The question was never whether AI could generate food.
Fully generated food was detectable, and it read as fake. We were already using AI on licensed stock, so generating the dish outright was the obvious next step — and the wrong one. Not because the output was bad, but because a customer who recognizes a synthetic dish stops believing everything else on the page.
The question I put to the team in March 2026 was not what the tool could do. It was where synthetic generation would break the promise we make to customers.
That set the constraint. Real food remains the visual truth layer. We shoot our own food; generation builds the scene around it.
A shoot designed to be reused.
There was already a budgeted April test-kitchen shoot on the calendar. Rather than commission something new, I asked whether that shoot was still necessary given what generation could now do — and then repurposed it as the capture engine.
One design decision made the output reusable: capture ingredients in several states rather than finished plated dishes. Culinary Ops supplied that rule. A finished dish photographs once and serves one recipe. A protein captured several ways becomes source material for many.
The plan was built on that principle — fifteen protein versions across chicken, pork, steak, and seafood, roughly eighty-five planned frames, about ten lifestyle and hands shots, seventeen cooked versions across four categories, neutral seasoning throughout so nothing committed the ingredient to one dish. Two cameras, pan and cooking-surface views, plated whole and sliced, three-quarter and overhead. No stylist. The point of that shot list is not how many frames came back; it is that reuse was designed into the shoot rather than discovered afterward.
I kept the scope deliberately small — a test shoot with a short list — and assigned a liaison to bridge Culinary rather than carrying that coordination myself. The pilot shot in one day, April 13, 2026, at the HQ test kitchen.
What is mine and what is not.
I originated the concept, named it, set the authenticity constraint, secured budget and scope, and framed the decision. The production method is not mine. The April production briefs — the photo team brief and the chef shot list — were authored by the practitioner who ran the shoot, along with the production itself, the collection, the composites, the SOP, and the later adaptation of the method to a second engine.
Stating that split precisely is not modesty. It is the actual proof.
The system outgrew me.
A reusable source collection went into the DAM on April 23. At the time of a September capture it held 150 source captures — the size of the source library at that moment, not a count of generated outputs. By June 2026 the outputs were in routine use in the Culinary production flow — described that way by the practitioner, in a message written at the time rather than in hindsight.
No record after April shows me producing a Virtual Pantry image. By September the practitioner was independently testing a vendor engine, working with its support team, identifying the expert-skill bottleneck, proposing a more scalable workflow, and reporting into the scaling conversation directly. A fourteen-page SOP exists that I did not write.
I stayed on the decisions that had to stay with me: joint gatekeeper for consumer-facing generated imagery; the rule that places our own captures in the cleared lane; and the indemnity and human-element questions escalated to Legal.
Where it actually stands.
This is an operational pilot and an emerging production system. Outputs reached routine Culinary use, and a second category — baked mixes — was routed into it rather than to an outside vendor. Executive sponsorship for wider scaling was still being pursued in September 2026.
The May 2026 presentation closed on the phrase that still describes the position: Enhanced, Not Replaced.
The durable part was never the model. It was the chain underneath it — real chef-prepared food, captured on purpose, held in a library, composed into a scene by the machine, and reviewed by a person before it reaches a customer. Better images came from better source material, not from a better prompt.
