The first attempt is cheaper, but a good final sometimes requires a second candidate.
In a seven-day slice, 317 of the generated generations were finalized. For 74 of them, the system created a repeat candidate. The estimated value of an image in the retry group was about 2.6 times higher, but comparing only the cost of trying is wrong: what matters is the price of a suitable result.
What exactly is counted?
We took the impersonal events of completed generations in seven days and divided them according to their characteristics. quality_gate_retry_countThere were 243 results in the non-retry group, 74 with a single repeat. The estimated cost of model calls for an image was taken into account; infrastructure, verification, storage and development are not included.
| Group | Results | Average image-cost | Median quality score |
|---|---|---|---|
| No retry. | 243 | $0,039 | 96 |
| With retry. | 74 | $0,103 | 95 |
Why the Cost of Trying Is Misleading
A cheap frame is useless if it doesn't fulfill a request. It's better to count the costs before the first suitable result: candidate generation, automatic evaluation, latency and the likelihood of manual reworking. Retry is more expensive, but can replace user re-starting and preserve the context of the task.
The average estimated cost of image generation among the outputs with a known number of repetitions.
Retry is justified if the cause is random and verifiable
An extra finger, torn edge or repeat texture may disappear in a new candidate.
Repeat is useful if the system amplifies a specifically ignored instruction.
The borderline option can be improved without changing the design.
Let’s assume one controlled repetition with enhanced personality preservation.
When Automatic Repeat Just Burns Your Budget
Retry does not create information that is not in the source, resolves a prompt contradiction or turns the generator into an accurate typography engine. If a person completely closes up a complex object, requires long text, or the user asks for five simultaneous transformations, you first need to change the task statement.
The useful limit is one automatic repeat for a diagnosed reason, and then it's better to show an obvious recommendation: choose another source, divide the task or name the object more accurately.
What to count as a product team
- The cost of a suitable final.All candidates and checks prior to issuance.
- Share of rescued retry.How many weak first attempts have been helpful.
- Delay p50 and p95.Quality doesn’t have to be an endless wait.
- Reason for replays.If one error dominates, it is more profitable to fix the prompt builder.
- Action after the result.Repeating, editing, paying or leaving shows real value better than score.
Three scenarios with different cost of error
Avatar. The loss of similarity makes a beautiful frame useless, so one retry is justified even with high technical assessment. Removal of the object. Repeat makes sense if the candidate left an artifact or cut a line; if a hidden background is unknown, a new generation will only offer another guess.
Creative illustration. Variation is permissible, and control should only intervene when there is a clear defect. For a product card, on the contrary, an altered packaging or incorrect label leads to complete alterations. The retry threshold should depend on intent, not be the same for any image.
Retry should buy quality, not mask system error
One diagnosed repeat is reasonable. Repeat for no reason is expensive and slow. The main metric is not the price of a call, but the cost of an outcome that a user is willing to save or buy.