ShotProbe

What one rejected AI shot costs

· original measurements

The figures in this article are listed in our data index, with what each one was measured from and when.

Some links on this site are affiliate links. If you sign up through them we may earn a commission at no extra cost to you.

A four-shot project cost ¥1.80 to generate. One shot was rejected in review. We regenerated that shot with an amended prompt, and the retry added ¥0.45 to the bill. The project totalled ¥2.25.

That makes the marginal cost of replacing one rejected shot ¥0.45, which is 25% of the original job cost and 20% of the final total. We reported this number first in the itemised bill for a 38-second ad; this article measures it in a controlled loop.

The number matters because it changes the question. When retries cost hours and conversations, teams avoid them. At ¥0.45 the constraint moves from budget to judgement: can you tell which shot needs fixing, and can you describe what is wrong clearly enough that the second attempt works?

The project

Four vertical shots, 5 seconds each, 768P. Generated separately, reviewed one by one, then stitched. The script and product were locked before generation, so this was assembly rather than exploration.

Initial generation cost ¥1.80. Review flagged one shot: the close-up was not tight enough, and a detail needed to read more clearly. We amended the prompt for that slot and regenerated only that shot. The regeneration added ¥0.45.

Final cost: ¥2.25 for four usable shots, one of which went through two attempts.

What the retry included

The ¥0.45 paid for:

It did not pay for:

Those four steps took about ten minutes of a person’s time. At ¥120 per hour, that is ¥20 in labour, more than forty times the generation cost. The generation is the smallest line on the real bill.

The review loop that produced the number

We ran this deliberately as a closed loop to measure what a rejection costs in practice:

  1. Generate four shots
  2. Review all four
  3. One rejected, with specific notes
  4. Amend the prompt for the rejected slot only
  5. Regenerate that slot
  6. Recheck the regenerated shot
  7. Assemble and deliver

The rejection was real. The shot genuinely did not meet the brief. This was not a stress test; it was a working project where one output needed fixing, and we measured the fix.

The loop worked because the system can regenerate a single slot without touching the others. That is a function of architecture rather than model capability: shots are generated independently and keyed by slot, so replacing one does not mean rerunning the whole job.

Why retries are cheap but mistakes are not

At ¥0.45 per retry, you can afford to regenerate freely. Five attempts at one slot cost ¥2.25, the same as the whole four-shot project. The generation budget stops being the bottleneck.

The bottleneck becomes knowing what to ask for. If the first attempt fails and the prompt stays the same, the second attempt will fail in the same way. The model does not learn from rejection; it runs the new prompt with the same process it ran the old one.

That makes the review note the most important input to the retry. “This does not work” gets you another random attempt. “The subject’s face is obscured; move the camera angle to show it directly” gets you a targeted fix.

Writing that kind of note is judgement work, and it does not scale the way generation does. A person can review twenty shots in an hour, but writing actionable notes for five rejected shots can take longer than reviewing all twenty.

What this does to quality control

When retries are cheap, the sensible strategy is to generate multiple versions of uncertain shots and keep the best one. That moves cost from generation to selection, and selection is slower work.

Consider two approaches:

Approach Generation cost Selection cost Total time
One attempt per slot, fix failures Lower Review happens once, then targeted retries Faster when yield is high
Three attempts per slot, pick best Higher Every slot needs comparative review Slower but more predictable

We use the first approach because our yield is above 70%, meaning most first attempts pass. If yield were lower, the second approach would make more sense: pay upfront in generation to avoid the back-and-forth of fixing individual failures.

The right threshold depends on two numbers: your yield rate, and how much a person’s time costs compared to a retry. At ¥0.45 per retry and ¥20 in labour per fix, you need a yield above 50% before single-attempt makes sense.

Where the real cost sits

The ¥0.45 is the itemised line. The real cost is in the three layers around it:

  1. Knowing a shot failed. You have to watch it, compare it to the brief, and decide. That takes time and judgement.
  2. Knowing why it failed. The prompt might be vague, the composition might be wrong, the motion might not fit. Identifying the cause determines whether the retry works.
  3. Knowing when to stop. Three failed attempts might mean the prompt needs rewriting from scratch, or it might mean the slot is asking for something the model cannot do.

None of those three steps get faster as the model gets cheaper. The ¥0.45 scales with model pricing. The ¥20 in labour per fix scales with how good your review process is, and that is a workflow problem rather than a pricing problem. This mirrors the structure we found in the cost of finished ads: generation is the smallest line, and the real work is in knowing what to generate and how to use it.

What this article does not claim

This is one project, one rejection, one model, one channel. It establishes that retries are cheap enough to use freely, not that every rejection will cost exactly ¥0.45.

Costs change when:

The labour side changes when your review process improves or when you are working in a category you know well enough to write precise prompts on the first attempt. Whether the output is worth using at all is a separate question: we measured that in a live account with 26,013 AI creatives.

We are measuring more projects as they go through review, and we will publish updates when the numbers change or when we find that the conclusion does not hold.

What we are doing about it

We are building two things to bring the labour cost down:

  1. Review templates by shot type. A-shots, B-shots and end cards fail in predictable ways. A checklist for each type makes review faster and notes more actionable.
  2. Prompt amendments that work. When a shot fails for a known reason, we have a library of amendments that fix it. That turns “write a better prompt” into “apply fix #3”.

Both are documentation work, not model work. The generation cost is already low. The next margin is in reducing the time it takes to catch a failure and describe a fix.

Method

This was a real review loop, not a simulation. The rejection was genuine, the retry solved it, and the project delivered.

← All articles