Diagnose and Fix Common Image Failures

The default response to a flawed generation is to press generate again. Sometimes that works, which is unfortunate, because it teaches a habit that fails on exactly the defects that matter. Rerolling solves one of four failure classes. For the other three it is a way of spending money to stay in the same place.

Four causes that look alike and are not

Sampling luck. The composition is fine, one detail landed badly, and a different seed would have avoided it. This is the class that rerolling fixes, and you can identify it by generating four samples: if three are clean, the fourth was luck.

A prompt that never said it. The image is a faithful rendering of an underspecified request. If four out of four samples get it wrong in the same way, the model is not failing — it is answering a different question. No seed will fix this, and rerolling twenty times is a way of avoiding rewriting one sentence.

A model limitation. Small text, precise counts, complex hand interactions, accurate logos, and legible interfaces are hard for most current image models in a way that is structural rather than random. Rerolling produces new variations of the same defect.

A settings problem. Duplicated heads and repeated architecture come from generating far above the model's native resolution. Plastic, over-smoothed skin comes from too much guidance. Mushy detail comes from too few steps. These are reproducible and reproducibly fixable.

The diagnostic is the four-sample test, and it takes one minute. Same prompt, four seeds. Count how many share the defect. One means luck. Four means the prompt or the settings.

Fix in cost order

Once you know the class, fix in the order that costs least, because the cheap fixes frequently make the expensive ones unnecessary.

Start with the seed if the four-sample test suggested luck — free, instant.

Then the prompt, which is the highest-leverage change available and the one most often skipped. Add the missing specification, and remove any negative phrasing that is not doing work. Note that describing what you do not want in the positive prompt frequently produces it; models attend to the noun, not the negation.

Then settings: resolution back to the model's native range, guidance down if the image looks over-baked, steps up if detail is mushy.

Then inpaint the specific area. Masking one hand and regenerating it at higher resolution is cheap, targeted, and does not risk the rest of the image.

Then post-process in an editor. There is no rule that says the fix must happen inside the model, and for a stray artifact a two-second clone-stamp beats twenty generations.

Only then change the model or add a LoRA, which is the most expensive move because it invalidates everything you have already tuned.

Some defects are the wrong tool, not a bad result

A few failure modes are worth recognising by sight because the correct response is to stop generating:

• Text in the image. Most image models cannot render more than a few words reliably. Generate the image without text and add the text in a layout tool. This is also how professional design has always worked.
• Exact object counts. "Five apples" is unreliable. Generate more than you need and remove, or compose.
• A specific real logo or brand mark. Beyond capability, this is usually a trademark problem too. Place the real asset in post.
• A named real person. Many services block it, and where they do not, publishing rights are a separate and harder question.
• Legible charts, interfaces, or documents. These will be plausible and wrong, which is worse than obviously wrong for anything a reader might believe.

Recognising these early saves the twenty-generation spiral. The image model produces the picture; another tool produces the precision.

Practice: build your own defect log

For two weeks, every time you discard a generation, write one line: what was wrong, which class it was, and what actually fixed it. Twenty lines is enough to see your own pattern, and the pattern is nearly always concentrated — most people find that a majority of their reruns come from two or three recurring specification gaps in how they write prompts.

Then turn those into a checklist you run before generating rather than after. Prevention is roughly ten times cheaper than repair, and the whole point of a defect log is to move work from the second to the first.

Common mistakes

Rerolling as a first response. Free-feeling, and it is not free — it costs time and credits, and it hides the real cause.

Piling up negatives. A long negative prompt is usually a sign that the positive prompt is underspecified. Fix the positive one first.

Fixing the wrong instance. Repairing the fifth-best image because it was closest is often slower than regenerating from a better base. Judge the base before you invest in it.

Not looking at full resolution. Hands, eyes, and text look acceptable in a thumbnail. Review at 100 percent before you decide an image is finished.