Cost Per Accepted Asset: Measuring Generative Production Properly

Every generative production pitch contains a per-unit number, and almost every one of them is measured at the wrong point. Cost per generation, cost per second, cost per credit: all of these count outputs. None of them count assets, and an output that failed review is not an asset. It is a cost.

The formula

That is the whole thing. Its power is entirely in what it refuses to exclude. A run that produced sixty frames and shipped four has a cost per accepted asset fifteen times its cost per generation, and the sixty is the number that left your account.

Worked example

Take a supplements brand wanting nine paid-social variants with one presenter, plus thirty product stills. Two batches, two very different profiles. The rates below are illustrative rather than a rate card, and the arithmetic is what matters.

The interesting line is not the total. It is that the presenter variants cost roughly three times per asset what the stills did, despite fewer renders, because the human hours are where the money went. Measured per generation you would have concluded the opposite: the two look almost identical, and the difference that matters disappears.

What the number tells you that nothing else does

• —Which shot types to stop accepting briefs for. If legible packaging type costs you £180 an asset and everything else costs £40, that is a pricing decision rather than a craft complaint.
• —When a pipeline change paid for itself. Build a trained identity instead of re-uploading references, watch the presenter acceptance rate move from 24 to 60 per cent, and the investment justifies itself in one campaign rather than in a slide.
• —Where the human time actually goes. Almost always review, almost never generation, which is the opposite of what the tooling narrative suggests.
• —Whether a job should be generative at all. When cost per accepted asset approaches what a photographer would have charged, the honest answer is to book the photographer.

The three ways it gets fudged

1. 01Excluding operator hours because they are salaried. Salaried time is still capacity, and capacity spent on regenerating a bottle is capacity not spent on the next brief.
2. 02Counting an asset as accepted when it shipped after two hours of retouching. It was not accepted, it was rescued. Log it as a failure plus a rescue cost, or the number quietly stops meaning anything.
3. 03Measuring per campaign rather than per shot type. An average across easy and hard work hides exactly the variance the metric exists to expose.

Running the ledger

This only works if the logging is automatic. Nobody fills in a spreadsheet at eleven at night after a render queue finishes. The pipeline has to write its own ledger: batch identifier, shot type, renders consumed, model spend, gate outcomes, operator time, final disposition.

Once that exists, a credit ceiling per asset becomes enforceable rather than aspirational. The run halts when a shot has consumed its allowance and asks a person to decide, instead of quietly spending past a number that was agreed in a meeting and never encoded anywhere.

A pipeline that cannot tell you what its last batch cost per accepted asset is not a pipeline. It is a habit.

What to ask a studio

Three questions, and the answers are diagnostic.

1. 01"What was your cost per accepted asset on the last job like this?" A studio running a ledger gives you a figure and a shot type. One that is not gives you a range for the whole engagement.
2. 02"What is your acceptance rate on shots with legible packaging type?" The honest answer in 2026 is a low fraction. Anyone claiming eighty per cent has either not tried it or is not checking properly.
3. 03"What is the credit ceiling per asset and what happens when a batch hits it?" It should halt and escalate. If nothing happens, there is no ceiling, only an intention.