Most bad AI images fail in a handful of predictable ways. This guide pairs each common problem with the reason it happens and the specific prompt change that usually fixes it, so you can stop re-rolling the same prompt and start correcting it.
- Most flaws come from too few pixels for a detail (hands, small faces, text) or from a prompt that gives the model conflicting or vague instructions.
- IMAGEKA picks one of several open-source models for each request, based on your style preset and wording. The Negative Prompt field works with the photo, anime and painterly models but is largely ignored by the fast general model, so describing what you do want is the fix that works everywhere.
- Give each subject its own sentence. That one habit fixes most swapped colours and merged details.
- There is no inpainting or editing in IMAGEKA. For a small flaw in an otherwise good image, a free image editor is faster than regenerating.
- Images are kept only in your browser. Download the ones you want to keep straight away.
Quick diagnosis table
Find your symptom, try the first fix, and jump to the section if it does not work.
| Symptom | Most likely cause | First thing to try |
|---|---|---|
| Extra or fused fingers | Hands are small and complex | Make the hands larger, or describe a simple pose |
| Warped faces in a group | Too few pixels per face | Fewer people, closer framing, 1024 px |
| Misspelled text | Long or unusual wording | Put 1–3 short words in quotes and say where they go |
| Colours or details swapped between subjects | Attribute bleeding | One sentence per subject, with positions |
| Wrong number of objects | Models count badly past about three | Describe an arrangement, not only a number |
| Head or feet cut off | Shot size not stated, wrong aspect ratio | Name the shot size and use 768×1024 for standing figures |
| Busy, messy image | Too many nouns, no background described | Cut the props and describe a simple background |
| Flat, dull lighting | No light source named | Name the source, its direction and the time of day |
| Plastic, waxy skin | "Perfect" wording and the model's own polish | Ask for natural texture and candid light |
| Random lettering or watermark-like marks | Words that suggest posters, covers or stock photos | Remove those words and describe the scene itself |
| Wrong style | Style preset conflicts with your prompt | Switch the preset to Default |
| Style different from what you expected | Your preset or words sent the request to a different model | Pick the preset that matches the look, or add a clear medium word |
| Blurred result | The safety checker flagged the output | Make clothing and context explicit |
| "Generation failed" | Wrong security code, a blocked prompt, a busy server or the daily limit | Check the code, reword the prompt, wait and retry |
| Images gone | Browser storage was cleared | Download every image you want to keep |
Extra fingers and bad hands
Why it happens. Hands have many joints, appear in countless poses in training photos, and often overlap with objects or each other. In a typical image they also take up very little space. Image models work on a compressed grid in which each unit stands for a block of output pixels, so a hand in a full-body shot might be only a few units across. That is not enough room to work out five separate fingers, and you get fused, doubled or extra ones. Current open models handle hands far better than early image generators, but none is immune. For the specialist models (photo, anime and painterly), IMAGEKA adds default negatives such as "bad hands" and "extra fingers"; that lowers the error rate but does not remove it.
Fixes.
- Give hands more pixels. Use 1024×1024 or 768×1024, and frame closer (a medium shot instead of a full-body shot). If you pick 512×512, the specialist models generate at 768×768 anyway, but the fast general model really does render at 512 px, where hands suffer most.
- Describe a simple, common pose. "Both hands wrapped around a coffee mug" or "hands resting in her lap" are poses the model has seen many times. "Gesturing expressively" invites chaos.
- Hide hands when they do not matter. Hands in pockets, arms folded, or cropped out of the frame.
- Do not ask for "perfect hands, five fingers". It rarely helps and wastes prompt space that could describe the pose.
a woman waving at a party, full body
Small, moving, open hand in a crowded scene: the worst case for finger errors.
Medium shot of a woman at a party holding a glass of sparkling water in her right hand, her left hand in her jacket pocket.
One visible hand in a familiar grip, shown larger in the frame.
Distorted faces in groups and small faces
Why it happens. It is the same pixel problem as hands. One face in a close-up gets thousands of pixels. Six faces in a wide group shot get a fraction of that each, so the eyes and teeth turn into smudges, and faces can blend into each other.
Fixes.
- Limit groups to two or three people when faces matter. For a crowd, accept that faces in the background will be soft and keep them out of focus on purpose: "a crowd softly blurred in the background".
- Use the largest resolution that fits the format, and frame from the waist up instead of head to toe.
- Give each person one or two distinguishing details (see attribute bleeding) so the model does not make them all the same face.
- Lens and depth-of-field wording helps: "shot at 50mm, the two people in sharp focus, background out of focus". The lighting and camera guide explains what these terms do.
Garbled or misspelled text
Why it happens. Image models draw letters as shapes, not by typing characters. Short, common words come out reliably. Long phrases, rare words and small lettering fall apart. Of the models IMAGEKA uses, the fast general-purpose model is by far the best at lettering, and putting words in quotes sends your request to it. The specialist models (photo, anime, painterly) are weak at text.
Fixes.
- Put the exact words in double quotes and keep them short: one to three words is reliable, a full sentence is not. The quotes matter twice: they mark the exact wording and they route the request to the model that handles text best.
- Say where the text goes and what it is printed on: "the word "OPEN" painted in white capital letters on the shop window".
- Describe the lettering style ("bold sans-serif capitals", "hand-painted script") so the letters get more of the model's attention.
- Generate four images and pick the one that is spelled correctly.
- For anything that must be right (prices, dates, a business name), leave clear space in the image and add the text yourself in a design tool. Real type is sharper, editable and always spelled correctly.
Details ignored or swapped (attribute bleeding)
Why it happens. You ask for "a man in a red hat and a woman in a blue shirt" and get a man in a blue hat, or both people in red. The text encoder turns your prompt into a set of meanings, and in a run-on list it is often unclear which adjective belongs to which noun. Colours are the worst offenders because they spread easily across the whole image. A long prompt with many attributes also means some simply get dropped.
Fixes.
- Write one sentence per subject, and give each one a position: "On the left… On the right…".
- Put the most important subject first. Early details tend to get the most weight.
- Cut attributes you do not care about. Every extra colour is another chance for bleeding.
- Use colours that are easy to tell apart, and do not repeat a colour word elsewhere in the prompt.
man red hat woman blue shirt sitting on bench park, yellow flowers
Tag-style list: the model has to guess which colour goes where, and "yellow" may spread to clothing too.
Two people sit on a park bench. On the left, an older man wears a red wool hat and a grey coat. On the right, a young woman wears a plain blue shirt. Yellow tulips grow in a flower bed behind the bench.
Each subject gets its own sentence and place, so the colours stay attached.
Wrong number of objects
Why it happens. Diffusion models do not count. They learned what "three apples" usually looks like, not how to place exactly three. One to three objects is usually right, and above about four it becomes guesswork.
Fixes.
- Describe a layout that implies the number: "three candles in a row, evenly spaced, the tallest in the middle".
- Spell out small numbers as words and repeat them through the layout: "two cups, one on each side of the teapot".
- For "many", say so honestly: "a bowl heaped with cherries" works better than "27 cherries".
- If the exact count matters (for example, a "5 tips" graphic), generate one object and repeat it yourself in an editor.
Subject cut off or badly cropped
Why it happens. If you do not name the framing, the model picks one, and in a square image a standing person often loses their feet or the top of their head. The aspect ratio matters as much as the words.
Fixes.
- Name the shot size: "full-body shot, the whole figure visible from head to shoes, space above the head".
- Match the format to the subject: 768×1024 portrait for standing people and tall buildings, 1024×768 landscape for wide scenes and groups.
- Ask for breathing room: "small in the frame, centred, with plenty of empty sky around it".
- If you plan to crop later for social media, generate wider than you need. The social media guide shows the crop maths.
Cluttered composition
Why it happens. Every noun in the prompt tends to show up somewhere. Ten props plus an unspecified background gives the model permission to fill every corner.
Fixes. Keep one subject and at most two supporting elements. Describe the background as plainly as the subject ("against a plain warm-grey wall", "on a clean white surface"). Ask for empty space by describing it: "the right half of the frame is open blue sky". Words such as "minimal" and "uncluttered" help a little. A described empty area helps more. The colour and composition guide covers layouts in detail.
Flat or boring lighting
Why it happens. Without lighting instructions you get the average of the training data: even, frontal, mid-day light. It is technically fine and visually forgettable.
Fixes. Name the light source, its direction and its quality. "Low evening sun from the left casting long shadows" or "a single window to the right, soft light, the far side of the face in shadow" changes the mood far more than "beautiful lighting" or "8K".
a cup of coffee on a table, professional photo, high quality
Quality words, but nothing about light, so the result is flat.
A cup of black coffee on a dark wooden table, lit by a single shaft of early morning sunlight from a window on the left, steam visible against the shadowed background.
A real source and direction give contrast, mood and visible steam.
Plastic, over-smoothed "AI look" skin
Why it happens. Models drift towards the polished look of retouched commercial photos. Words such as "flawless", "perfect", "beautiful" and "ultra sharp" push further in that direction, and so does the 3D Render preset ("subsurface scattering, octane render"). Portraits usually go to IMAGEKA's photorealism model, which renders pores and fine skin detail well when you ask for them. If your request lands on the fast general model instead (for example with the Default preset and no photo words, or when a specialist fails and the general model steps in), expect smoother skin and very strong background blur.
Fixes.
- Drop the flattery words and describe texture: "natural skin texture with visible pores and faint freckles, light laugh lines".
- Make the setting candid: "a candid snapshot", "everyday clothes", "overcast daylight". Soft, even daylight hides less and looks more real than studio glamour lighting.
- Try the Vintage preset. Film-stock wording adds grain and warmer tones that hide the waxy look.
- Ask for less blur if the background looks fake: "background in focus" or "shot at f/8".
Unwanted lettering or watermark-like marks
Why it happens. Many training images carried captions, logos and watermarks. Prompt words linked to those images bring the marks with them: "poster", "book cover", "magazine", "stock photo", "advertisement", "meme", "infographic", "label".
Fixes. Describe the image itself instead of the kind of media it would appear in: say "a dramatic mountain landscape at dusk" rather than "a travel poster of mountains". If lettering keeps appearing on signs or clothing, describe those surfaces as plain ("a plain white T-shirt", "a blank wooden sign"). Typing "no text, no watermark" in the Negative Prompt field helps with the photo, anime and painterly models (which already get "watermark" and "text" as default negatives), but the fast general model largely ignores it. If a mark is small, remove it in an image editor (see fixing small flaws).
Style preset fighting the prompt
Why it happens. Each preset simply adds text to the end of your prompt. Photorealistic is on by default and adds "photo taken on Sony A7R, 85mm lens, shallow depth of field, photorealistic, ultra sharp". If your prompt says "a loose watercolour of a harbour", the model gets two opposite instructions and you get a muddy hybrid.
Fixes. When your prompt describes a style, set the preset to Default. When you use a preset, leave style words out of your prompt and describe only the content. The art styles guide lists exactly what each preset adds.
The style came out different from what I expected
Why it happens. IMAGEKA does not use one model. For every request it picks one of several open-source models from your preset and your words: a fast general-purpose model (the default, also used for text in quotes, logos, posters and icons), a photorealism specialist (realistic people, portraits, products and food), an anime/illustration specialist and a painterly/traditional-media model. Each has its own look, so a small change in wording can switch models. "A woman reading in a café" with the Photorealistic preset goes to the photo model; add the word "poster" and it goes to the general model, which looks noticeably different. You cannot choose the model directly, but you can steer it.
Fixes.
- For a realistic photo of a person, product or food, keep the Photorealistic or Vintage preset, or with Default add a photo word such as "photograph" or "close-up".
- For anime or manga, pick the Anime preset or use the word "anime". That model also understands short comma-separated tags ("1girl, silver hair, school uniform").
- For paint, pencil or pixel looks, pick Oil Painting, Watercolor, Sketch or Pixel Art, or name the medium ("watercolour", "charcoal", "pixel art").
- Watch for words that pull towards the general model: anything in quotes, and words like "logo", "poster", "icon", "label" or "vector". They win over a style preset, because that model handles text and graphic layouts best.
- If a run fails on a specialist model, IMAGEKA automatically retries with the general model, so an occasional image may look different from the rest of a series. Running the same prompt again usually brings the expected style back.
Images too similar, or too random
Too similar. IMAGEKA uses a new random seed on every run, but the fast general model in particular keeps close to its strongest reading of your words, so four images from the same detailed prompt can look almost identical. Re-rolling will not change much. Change the prompt instead: a new camera angle, time of day, setting or colour palette. Leaving one dimension open on purpose ("in an unusual setting") also adds variety.
Too random. If every run looks completely different, the prompt is under-specified and the model is filling the gaps differently each time. Pin down the subject, framing, light and palette. IMAGEKA has no seed control, so for a consistent series keep one fixed "anchor" phrase for the style and lighting, and change only the subject.
Blurred output from the safety checker
Why it happens. After an image is generated, an automatic safety checker looks at it. If it judges the image to be possibly explicit, it blurs it. Like every classifier it makes mistakes. Innocent prompts sometimes produce an image the checker reads as revealing, for example beach scenes, swimwear, bathing, people in bed, medical or anatomy subjects, classical sculpture, or close-ups where large areas of skin fill the frame.
Fixes for innocent prompts.
- State clothing explicitly: "wearing a long-sleeved rash guard and board shorts" rather than "at the beach".
- Widen the shot so skin is not filling the frame, and add context: "a family having a picnic on the sand, umbrellas and beach bags around them".
- Swap ambiguous words: "relaxing on a sofa with a book" instead of "lying in bed".
- For art references, describe the subject instead of naming nude works or poses.
These tips are for prompts that were flagged by mistake. A separate content filter blocks prompts involving sexual content with minors, extreme violence and similar material, and those will not generate however they are worded.
"Generation failed", security code errors and the daily limit
- Wrong or expired security code. Each generation needs the 5-digit code shown in the security image. A code works only once and expires after about three minutes, so if you wrote a long prompt after the code loaded, or are generating a second time, click the refresh icon for a new code and type it carefully.
- Blocked prompt. If the content filter rejects the prompt, rewording the same idea will not get round it. If you think it was blocked by mistake, rephrase the ambiguous words as described above.
- Busy or slow server. Generation goes through an external inference service. Most images arrive within seconds and occasionally take up to about 30 seconds. If a request times out, wait a minute and try again rather than clicking repeatedly.
- Daily limit. To keep the tool free there is a fair-use limit of 100 generations per day per IP address. When you reach it you will see a "limit reached" message, and the count resets at the start of the next day. Offices, schools and some mobile networks put many people behind one IP address, so you may share the allowance without knowing it.
Each run counts as one generation, whether it makes one image or four. Once you are close to the prompt you want, ask for 4 images per run to see more variations for the same share of the allowance. Failed requests can also count, so if something goes wrong wait a moment before retrying rather than clicking Generate again and again. (Failures caused by a problem on our side do not count.)
Images disappeared
IMAGEKA has no accounts and does not keep a gallery on its servers. Your recent images are saved in your own browser's local storage, which is why they can vanish. Common causes:
- You cleared browsing data or cookies, or a cleaner app or privacy extension did it for you.
- You used a private or incognito window, which throws away storage when it closes.
- You opened the site in a different browser or on another device. Storage is never shared between them.
- The browser ran out of storage space or cleared old site data on its own, which some mobile browsers do.
Once storage is cleared, the images cannot be recovered. Download every image you want to keep as soon as you see it.
Fixing small flaws without regenerating
IMAGEKA has no inpainting, image-to-image or editing tools. That is an honest limitation. If an image is 95% right but has a sixth finger, a stray letter or a smudge on a face, regenerating may lose everything you liked about it. A free image editor is usually quicker:
- Healing or clone tools (in GIMP, Krita or the browser-based Photopea) remove stray marks, watermark-like smudges and small artefacts in a few seconds.
- Cropping removes a bad hand or a cut-off object at the edge of the frame.
- Adding real text in a design tool replaces garbled AI lettering. Paint over the bad lettering first.
- A light blur or grain over a waxy face can make it look more photographic.
For the underlying habits that prevent most of these problems in the first place, see how to write prompts that work, and for what terms such as seed and guidance mean, see the glossary.
Frequently asked questions
Why does the Negative Prompt field not remove things?
It depends which model handled your request. The photo, anime and painterly models do use the negative prompt (and IMAGEKA adds sensible defaults such as "bad hands", "blurry" and "watermark" for them). The fast general model, used by default and for text, logos and posters, is built for speed and largely ignores it. Either way, describing what you want works more reliably: say "a plain blue sky" rather than putting "clouds" in the negative field.
Why did one image in my series look different from the others?
Small wording changes can send a request to a different model, and if a specialist model fails, IMAGEKA retries with the general model. Keep the preset and the style words identical across a series, and regenerate any image that looks out of place.
Will generating the same prompt again fix a bad hand?
Sometimes, because each run uses a new random seed. If the problem keeps coming back, the cause is in the prompt or the framing. Make the hands larger, give them a simple pose, or hide them.
Does adding "high quality, 8K, masterpiece" help?
Very little. Those tags mattered more for early image models, and IMAGEKA already adds suitable quality defaults behind the scenes where a model benefits from them. Words about light, framing and texture improve the image far more.
Why are some of my images blurred?
The safety checker blurs outputs it thinks may be explicit. If your prompt was innocent, say explicitly what people are wearing, widen the shot and add context, then generate again.
Can I get back images I generated yesterday?
Only if they are still in the same browser on the same device and its storage has not been cleared. IMAGEKA does not keep copies on its servers, so download anything you want to keep.
Is there a way to edit just one part of an image?
Not in IMAGEKA. There is no inpainting. Use a free image editor for small fixes, or change the prompt and generate a new image for bigger changes.
