Troubleshooting: Twelve Reasons an AI Generation Goes Wrong

Generation problems tend to look mysterious and turn out to be mundane. This is a symptom-first reference: find the thing that went wrong, read the cause, apply the fix. Most of these take one attempt to resolve.

Upload and input problems

1. The upload is rejected. UPretty accepts JPG, PNG and WebP. The formats that get rejected in practice are HEIC files straight from an iPhone, RAW files from a camera, and PDFs. Convert to JPG first — on an iPhone, setting Camera to "Most Compatible" makes this permanent.

2. Nothing happens after selecting a photo. Usually a very large file on a slow connection. A 40-megapixel photo is not more useful than a 12-megapixel one for this purpose; resize the long edge to around 2000 pixels before uploading.

3. The face is not detected. The face is too small in the frame, too heavily obscured, or at too extreme an angle. Crop closer so the head fills a meaningful portion of the frame, remove sunglasses and masks, and shoot within thirty degrees of straight on.

Identity problems

4. The result does not look like you. The most common failure. Causes in order of frequency: a soft or compressed source photo, overhead or backlit lighting, an extreme angle, a heavily stylised theme, or a filtered source image. Regenerate once first, then work through those in order.

5. The age looks wrong. Almost always the camera angle. Shooting from below widens the jaw and reads as older or heavier; shooting from above enlarges the eyes and reads as younger. Hold the camera at eye level.

6. Your glasses changed shape. Eyewear is reconstructed rather than copied, and thin frames are the least reliable element in the whole pipeline. Either accept variation and generate a few times, or use a source photo without glasses.

Composition problems

7. The top of the head is cut off. Your source photo was cropped too tightly, so the model had no hairline to work from and improvised the top of the composition. Leave a margin above the hair in the source.

8. Hands look wrong. The known weak point of image models. Fingers merge, multiply or bend incorrectly. The reliable answer is to use a source photo with hands out of frame and choose themes that do not call for visible hands.

9. Text in the image is gibberish. Expected behaviour, not a bug. Generated lettering on signs, badges and clothing is decorative noise. Never use a generated image where the text has to be readable.

10. The framing is wrong for where you are posting it. Aspect ratio was chosen for a different destination. Decide the destination first — square for profile pictures, 9:16 for Stories, 16:9 for thumbnails — and generate to that shape rather than cropping afterwards.

Process problems

11. The generation failed partway through. Transient errors happen. A failed generation should not count against your daily allowance; if your count dropped without an image being delivered, contact support with the approximate time and it can be checked.

12. Your image disappeared from the gallery. Generated images are retained for thirty days and then permanently deleted. There is no recovery after that window. The habit worth building is downloading anything you like on the day you make it.

The three fixes that solve most cases

If you take nothing else from this list: regenerate before diagnosing, because randomness accounts for a large share of bad results. Replace the source photo with one that is sharp, front-lit and front-facing, because input quality caps output quality. And match the theme to your goal — realistic themes for recognisable portraits, stylised themes for everything else.

When it is not you

Some limitations are inherent and no amount of troubleshooting removes them: distinctive scars, birthmarks and tattoos are often softened or dropped; very curly and very fine hair are both reconstructed approximately; and images containing more than one person preserve identity much less reliably than single-subject portraits. If you hit one of these, the answer is to work around it rather than to keep adjusting inputs.