Your GPU, your project Crypto payment without KYCHow to pay
English
My Account
First project guide

A useful trial also ends with a decision to stop.

Continue a trial if the next run answers a specific question, can be checked, and fits within the time remaining. Modify it if an identifiable cause deserves a bounded fix. Stop if the essential quality remains out of reach, if the trials repeat the same failure, or if collecting the results is likely to be sacrificed. The amount already spent explains your track record; it is not enough to justify extra work without a clear prospect.

On this page

1. Write down what would make the trial succeed or end

Before starting a series, note the expected result, the flaws that make the output unusable, and the moment of the decision. A goal like "twenty-four images accepted in the requested format" is easier to check than "make beautiful images." A learning goal can also be valid: completing a small run, finding the settings again, and checking a copy.

Set a personal time limit and a cap for the extra spending you are considering. They can be modest: a single correction hypothesis before the review, then a save. These are not universal values. Choose them based on your task and the person who will approve the result.

Also set the criterion that cannot be negotiated at the last minute. A response that invents a procedure, or a visual whose product is distorted, can remain rejected even if everything else seems convincing. Otherwise, the fatigue of the trial may gradually turn a blocking flaw into an acceptable detail.

2. Work out the time actually available for another attempt

Start from the time left in your session and in the usable rental window. Take the nearest constraint, then subtract the time planned for checking and copying the results. The balance can accommodate another attempt, its check, and the essential corrections. Do not count the same slot for both producing and saving.

A limit configured in software does not replace this organization. For example, Transformers notes that max_time can let the current generation pass finish after the specified time. This setting also does not reserve time for review or transfer. Keep your own stopping point and use the stop commands provided by your tool.

If the duration of the next run is unknown, reduce its scope: one input, one question, one export. If even this small trial cannot be checked before the deadline, postpone it. An output that appears just before the end but is never opened is not an accepted result.

3. Tell the three possible decisions apart

The difference between continuing and modifying comes down to what you have learned. Continue when the method meets the criteria on the checked scope and you are cautiously expanding the batch. Modify when you can name a cause and the change that tests it. Stop when you no longer have a useful question to resolve within your limits.

A repeated technical error calls for a diagnosis, not a series of unrelated new settings. An output that is technically complete but of poor quality calls for another review. Describe the failure in your own words before deciding. Buying more capacity does not answer a missing document or an instruction that asks for two contradictory things.

3. Tell the three possible decisions apart
DécisionCondition utileProchaine action bornée
ContinueCriteria met on a relevant sample; checking time availableExpand a short slice with the same settings
ModifyPlausible cause, isolated change, and verifiable resultTest a difficult input, then an already successful case
StopLimit reached, no clear hypothesis, or a key result out of reachSave, note the blocker, and prepare another approach

4. Keep the package you committed to separate from the cost of the follow-up

The price of a BriefGPU package corresponds to one lot during the chosen period. In your review, keep that full amount multiplied by the number of lots. Do not replace it with a per-minute rate calculated after the fact, and do not turn unused time into assumed credit. The work decision does not create a new billing rule.

You can relate the package cost to accepted results, as long as you state the number and scope. If no result is accepted, the ratio cannot be calculated; showing zero would give the misleading impression of a free result. Rejected outputs stay in the trial review, but not among accepted deliverables.

Look at the follow-up separately: human time, known external costs, any other period to consider, and the additional result you hope for. Costs that are not documented remain unknown, not zero. The budget calculator helps you keep that distinction. A past expense does not prove that the next attempt will be useful.

Example: eighteen images accepted out of a goal of twenty-four

Let's take a teaching scenario with an RTX A5000 lot over three days, priced at 23.57 USD from the September 24, 2026 catalog. The freelancer's goal is to produce twenty-four accepted visuals. In this fictional scenario, eighteen pass the checks and six have a recurring flaw. The numbers illustrate a decision; they do not measure this GPU's output.

The planned ratio was 23.57 ÷ 24, or about 0.98 USD per result. With eighteen accepted results, the package ratio is 23.57 ÷ 18, or about 1.31 USD. External costs are not documented: no overall cost per result is stated. The gap between the two ratios describes the review, not a sufficient reason to rerun the six files.

Fifty minutes remain before the chosen session limit. The freelancer sets aside twenty minutes to copy and check the folder; thirty minutes remain for one attempt and its review. Their estimate for the new run is forty minutes, to which they add ten minutes of checking. This fifty-minute attempt does not fit in the thirty-minute slot.

So they stop production to keep the eighteen outputs and the six reasons for rejection. The goal of twenty-four is not declared met. If a partial delivery works for the recipient, that must be agreed explicitly. Another session can address a specific hypothesis; the current review does not require either resuming immediately or buying another period.

Example: eighteen images accepted out of a goal of twenty-four
Repère du scénarioCalcul ou étatConséquence
Time until the limit50 minutesStarting point
Copy and check set aside20 minutesTo preserve
Time for one attempt and its review50 − 20 = 30 minutesCeiling for additional work
Estimated new attempt40 + 10 = 50 minutesDoes not fit in the slot
Quality review18 accepted; 6 rejectedGoal of 24 not met

5. Stop cleanly and make the decision reusable

When the tool allows it, stop adding new tasks, then let the current item finish before closing. Follow its shutdown procedure. If the interruption leaves a file incomplete, mark it for review and keep it separate from accepted outputs. Closing a window or losing a connection is not a check of the work's state.

Keep a note of a few lines: goal, result obtained, limit encountered, decision, and condition for resuming. For example: "Eighteen outputs accepted; six details unreadable; production stopped to preserve the copy; resume after a targeted test on those details." Attach the settings and the list of relevant IDs.

Stopping your software trial does not, by itself, constitute a request for cancellation, refund, or order modification. For a question about the rental record, see the terms and the support process. This guide organizes your work and assumes no automatic commercial change.

The mistakes that stretch out a trial without improving the decision

Avoid pushing back the deadline after every failure, choosing only the best outputs for the review, or counting an unchecked output as accepted. Don't quietly lower the bar to make the result match the goal. If the need changes, write down the new goal: that's a project decision, not retroactive success.

Don't keep going just because time has already been spent on the work. Ask what the next attempt will let you learn or deliver, and with what check. A vague answer like "maybe it'll work this time" calls for a more precise hypothesis. When the best move is to prepare the files on your own machine, that preparation can come before another rental.

Frequently asked questions

Is stopping after a single failure always too soon?

No. A single failure can reveal an essential incompatibility or a missing input that you can't resolve during the session. On the other hand, a simple, well-understood defect may justify one more small trial. The decision depends on the cause and the limits, not on a set number of attempts.

Should I keep going if the plan is already paid for?

Not for that reason alone. Check that the next piece of work delivers a useful result or piece of information and leaves time to verify the outputs. The amount already spent stays in the review, even if you choose to spend the rest of the session on backups.

Can a trial with no deliverable still have been useful?

Yes, if its purpose was to test a hypothesis and if you keep a usable conclusion. A precisely described compatibility problem or an observed quality limit can keep you from repeating the same preparation. Don't present that learning as a deliverable you didn't produce, though.

Can I add another plan to finish later?

The calculator doesn't automatically combine several plans and doesn't assume any extension. If the project needs another period, prepare a new schedule and check the terms that apply to the job. First save whatever lets you pick the work back up.

Go at your own pace

A little method goes a long way at the start.

Open the guides