Most often, it is not the picture that breaks, but the connection between action and limitations.
The "do beautiful" request leaves too much freedom for the model, but a long five-act instruction doesn't work well: requirements start competing; it's more reliable to set one main action, name the goal precisely, and list separately what should remain unchanged.
How we calculated and what the numbers mean
The sample included 396 unique TrendMeBot generations in a seven-day slice, and for them the system performed 513 checks: some of the images were re-evaluated after automatic adjustment; we did not analyze user identities, private photo content and personal data — only anonymized technical features of a result.
Verification compares the image with the original prompt and reference in four directions: execution of instructions, preservation of identity and scene, visual defects and fitness of the result for use. One frame can get several causes at once, so the sums of individual errors are not equal to the number of generations.
What mistakes have been encountered most often?
The main signal is instruction_ignored: the model made a believable image, but did not perform the main action. no_transformWhen the result is too close to the source, then there's a prompt mismatch and problems physically building the scene.
The number of unique generations where the cause has met in at least one candidate; the causes may overlap.
The action is worded vaguely
"Improving," "do fine" and "process beautifully" don't set a measurable change. The model chooses the safe minimum: it changes light slightly or returns almost the same frame.
In one prompt, several competing tasks
Replacing backgrounds, clothing, posture, text and face simultaneously turns precise editing into a new generation. The more independent surgeries there are, the greater the chance of losing at least one.
Unnamed exact target
The phrase "get the extra" is understandable to a person from context, but models have to guess. It's better to name an object, its position, color or role in a frame.
It is not specified that to maintain
When the background is replaced, the model can change clothes, angles or facial features at the same time. A short list of invariants keeps the result closer to the source.
The source does not contain enough information.
If an object covers half of your face, or if you have to remove a person from a complex patterned background, the hidden area has to be invented.
Query formula: action → goal → preservation → criterion of the result
The work prompt doesn't have to be long. Its job is to remove four kinds of ambiguity. First, the verb tells you what to do. Then the goal shows you exactly what it's with. The constraints hold important parts of the source and the outcome criterion explains what the finished work looks like.
verb
location
changeable
sign
Remove the person on the left. You don't change the main character, your clothes, posture and lighting. Naturally restore the wall and shadow behind the remote person.
This formula is especially useful for removing objects, replacing backgrounds, improving quality, local retouching and writing. For a creative portrait, limitations can be relaxed, but the main task should still remain one.
How to rewrite a request without turning it into a technical task
"Enhance the photo."
→Testable."Improving facial sharpness, leveling the exposure and removing color noise. Face, clothing, background and framing don't change."
Change the background, clothes, posture, add text and make it as a cover.
→Step by step."First, just replace the background with a quiet studio. Save your man, clothes and posture."
"Take it all away."
→Exact target."Remove the red jar and the white wire from the table. Save all other objects and shadows."
“Make a different image, but don’t change anything.”
→Priority"Save face and pose. Clothes and background can be changed to a fantasy portrait."
Why One Beautiful Candidate Is Not Enough
A person quickly notices an obvious defect, but the automatic system has to distinguish between two different failures: one is a technically bad image: broken hands, face or perspective. The second is more complicated: the picture looks convincing but answers the wrong prompt. Both are equally useless for the user.
353 out of 396 Generations ended with a candidate who passed the final test.
117 generation It doesn't mean that all the first results were terrible: sometimes retry is necessary because of an unfulfilled action.
Automatic verification does not replace the taste of the user, but cuts off options that should not be given out as a finished work.
Checklist before generation
- Leave one main action.If you need background and clothing, change the background first, then edit the result.
- Name the target so that you can point it with your finger.Not “extra,” but a red jar to the right of the laptop.
- Identify 2-4 important elements.Face, clothing, pose, background, logo or location of objects.
- Describe a sign of readiness.The object has disappeared, the background is restored, face is recognizable, inscription reads without unnecessary letters.
- Evaluate the source.Severe darkness, overlaps and fine details increase the volume of what models have to invent.
- Do not confuse aesthetic quality with the task.First check the action and preservation of important details, then light and beauty.
The best request is not like a spell, but a short task statement.
The AI doesn't have to understand the hidden intent. The clearer the action, purpose and boundaries of change, the easier it is to distinguish a good result from just a beautiful one. And if the task is complex, a sequence of two precise edits is usually more reliable than one overloaded generation.