When an AI Generation Fails or Comes Out Wrong

Most generations work. The ones that do not fall into a small number of categories, and almost all of them are cheaper to fix than they look — because in nearly every failure case you were not charged.

This page covers what each outcome means, whether it cost you anything, and what to do. It is the reference for the first week, when everything is unfamiliar and every red status looks like a problem with your account.

First: were you charged?

No, unless something was produced.

Credits are held when a generation starts and only actually spent once a result is saved. If it fails, is cancelled or is rejected, the hold is released and the amount returns to your balance.

This is why a balance can briefly read lower than expected while work is running — that amount is reserved, not gone. For a batch it applies item by item: you pay for the items that produced something and not for the ones that did not.

If a balance still looks wrong an hour later, that is worth raising with support with a rough time.

The statuses

Status	What it means	Charged?
Pending	Accepted, not started	Held
Queued	Waiting its turn	Held
Processing	Being made now	Held
Completed	Finished, in your library	Yes
Partial success	Part of a batch worked	For the parts that worked
Failed	Something broke; nothing produced	No
Cancelled	Stopped, by you or the system	No
Rejected	Refused — usually a content rule	No

Failed

Something broke, and it is usually temporary. Work through these in order:

Try again. A retry succeeds more often than not, because the common causes are transient.
Try a different model. If one model keeps failing on a prompt that works elsewhere, the model is the problem rather than the prompt.
Simplify. Very large sizes, unusual combinations of settings and long prompts all raise the chance of failure.
If it keeps failing, contact support with roughly when you tried and what you were making.

Rejected

The request was refused rather than attempted. Nearly always the content rules: something in the prompt, or in an image you supplied, is not permitted.

Rewrite and try again. If you believe a rejection is wrong — an ordinary word that happens to appear on a list — raise it with support and quote what you were trying to make. That is how false positives get fixed.

Rejections do not cost credits.

If rejections are frequent rather than occasional, the issue is likely to be subject matter rather than phrasing. See rights, likeness and consent — likeness of real people is the most common cause.

Cancelled

Either you stopped it, or the system did — most often because it exceeded the time allowed. Long video generations at large sizes are the usual case, which is a good reason to generate at a normal size and upscale the keeper.

Try a smaller size, a shorter clip, or a faster model.

Stuck in Queued or Processing

The queue is busy. Everyone’s work is interleaved so that no single account monopolises capacity, which means waiting when demand is high. Higher plans get better queue priority, which is a large part of what they are for.

You do not need to stay on the page. Leave, come back, and it will be in your library. Turn notifications on if you would rather be told.

If something has genuinely been sitting far longer than that kind of work usually takes, raise it. Orphaned work is detected and its held credits returned, so a stuck job does not quietly cost you anything.

Completed, but the result is wrong

Not a failure. The system did what was asked; the request was the problem. This is the most common outcome in week one and the most fixable.

The quick checks:

Is what you want actually in the prompt, or only in your head? The most frequent cause by a wide margin.
Are you fighting a known weakness? Text in the image, exact counts, hands doing fiddly things, precise spatial arrangements, a specific real person’s face. These do not improve with more attempts.
Starting from an image — is strength too high or too low? Nothing changed means too low; your subject vanished means too high.
Did you say what the camera does? Unspecified camera is the single most common reason a clip reads as amateur. See prompting for video.

The video-specific failures

Symptom	Cause	Fix
Too much camera movement	Camera unspecified	Add “camera holds steady”
Subject warps or melts partway	Motion too complex	Ask for one simpler action
Nothing much happens	Description, not direction	Name the action explicitly
Clip shorter than expected	Length is model-dependent	Check what that model supports
Result looks nothing like the idea	Generated from a description	Start from an approved still
Face changes between shots	No locked character	Use an avatar
Published video not in the public feed	Content review not finished	Wait; see below

Published but not appearing

Not a failure either. Every published item is content-rated. Images are rated at once; video is reviewed after publishing, because the check samples the actual clip rather than a single frame — a video whose first frame is harmless is not necessarily harmless.

So a newly published video is in your own library straight away and in the public feed once its review finishes. If a rating is uncertain it stays pending rather than going public: unreviewed is not treated as approved.

This is not a queue you can jump, and nothing is wrong when it happens.

The failure that is not a failure: running out

Generating stops. Nothing is deleted, nothing is lost, and your library stays exactly as it is — you simply cannot start new work until you have credits again.

Two ways forward: top up, or move to a plan with a larger allowance. Before doing either, read what an episode actually costs — running out early is usually a drafting-habit problem rather than a plan problem, and the habits are worth more than the tier.

A triage order for week one

When something is wrong and you do not know where to start:

Check the status. It tells you whether this is a system problem or a prompt problem, and they have nothing in common.
If Failed or Cancelled — retry once, then simplify.
If Rejected — the content rules, not you. Rewrite.
If Completed and wrong — the prompt. Change one thing and run again.
If Queued for a long time — leave the page and come back.
If the balance looks wrong — check for holds on running work first.

Almost everything resolves at step two or step four.