Image Prompt Not Working? 20 Problems, 20 Fixes

TL;DR: An image prompt not working is three different problems wearing one costume. Some are wording problems a rewrite fixes. Some are parameter problems where the sentence is irrelevant and a request field is wrong. And seven of the twenty below are capability limits, where no phrasing helps and you need a different tool or an extra step.

Why is my image prompt not working?

Because you are probably treating every bad image as a writing failure, and about a third of them are not. Image prompt troubleshooting goes wrong when the diagnosis is skipped: you rewrite the sentence eleven times against a problem that was never in the sentence.

Three buckets, three different reactions.

Prompt problems. The model did what you said, and what you said was vague, contradictory or overloaded. Rewording fixes these, often on the first try.

Parameter problems. Your wording is fine. A field in the request is wrong, or the interface never exposed the control at all. Rewriting is wasted effort, and it is the most common wasted effort in this whole category.

Capability limits. The model cannot do it. Not "cannot do it yet with your phrasing" but cannot do it. Vendors say so themselves. Under its own Limitations heading for the GPT Image models, OpenAI writes that "the model can still struggle with precise text placement and clarity" and that "the model may have difficulty placing elements precisely in structured or layout-sensitive compositions" (OpenAI image generation guide, read August 28, 2026). Those are not prompt bugs. They are the shape of the tool.

Any guide that hands you a rewrite for all twenty is lying to you.

Which of the 20 problems do you have?

Find your symptom, note the number, jump to it. The third column is the part most guides skip.

Problems 1 to 8: the ones a rewrite really does fix

These are the good news. Each pair below shows the broken prompt and the corrected one, so the change is visible rather than described.

1. "It ignored half of what I wrote"

What is happening. Attention is finite and spread across the whole prompt. A fifteen-clause paragraph does not give each clause a fifteenth of the weight; the survivors tend to be concrete, early, and not competing with a neighbour. Adding detail to fix a missed detail usually makes it worse. Google's Gemini image guide gives the structural fix: "For complex scenes with many elements, break your prompt into steps" (Gemini API image generation, read August 28, 2026).

Fix. Cut to the three things you would be annoyed to lose. If the scene needs more, build it across turns rather than in one block.

Broken:
A cosy Scandinavian kitchen with white oak cabinets and brass handles and a
marble island and a copper kettle and hanging plants and morning light and a
ceramic bowl of lemons and a linen runner and a small dog by the door and a
wall clock and open shelving with mismatched mugs, photorealistic, 8k, detailed

Fixed:
A cosy Scandinavian kitchen. White oak cabinets, a marble island, morning light
from a window on the left. One copper kettle on the hob. Photographic, natural
colour.

2. "The style bled onto everything"

What is happening. Style words are global by default. You wrote "watercolour" meaning the poster on the wall, and the model applied it to the room, the person and the cat. There is no syntax in a plain prompt that scopes an adjective to one noun.

Fix. Decide whether the style is the whole image or one object. If it is one object, describe that object as a physical thing inside a photograph rather than as a style. If instead an attribute landed on the wrong subject, that is a different failure with a different fix, covered in why your image prompt produced the wrong subject.

Broken:
A woman in a modern office with a watercolour poster of mountains on the wall,
watercolour style

Fixed:
Photograph of a woman in a modern office. On the wall behind her hangs a framed
watercolour painting of mountains, soft washes of blue and grey. The rest of the
scene is a straight photograph.

3. "Colours are washed out or just wrong"

What is happening. Colour names are broad: "blue" spans navy to cyan. Meanwhile a stylisation setting or a film reference you added is quietly desaturating everything.

Fix. Name colours as materials and pigments, which carry strong visual associations, and say the saturation you want out loud. Pasting a hex value is my own habit rather than anything a vendor documents as supported, and I treat it as a hint, never a specification.

Broken:
A product shot of a bottle, blue, #1B4D8C, vibrant

Fixed:
Product photograph of a glass bottle in deep cobalt blue, the colour of ink,
against a warm sand background. Rich saturation, no colour cast, clean white
highlights.

4. "Light is coming from the wrong side"

What is happening. You described a mood rather than a direction. "Dramatic lighting" and "moody" tell the model nothing about where the key light sits, so it picks a plausible answer, usually a soft front-lit look.

Fix. Use the real vocabulary. Direction, quality, source, in that order. The 40 lighting terms in our lighting vocabulary reference are the words that actually change the render, and this post will not re-teach them.

Broken:
Portrait of a chef in a kitchen, dramatic moody lighting

Fixed:
Portrait of a chef in a kitchen. Single hard key light from camera left at 45
degrees, deep shadow on the right side of the face, no fill. Warm tungsten
practical in the background.

5. "It has that glossy AI look"

What is happening. Default aesthetics. Most models lean toward high contrast, high saturation and unnaturally clean surfaces, because that reads as impressive in a grid of thumbnails. Words like "8k", "hyperrealistic" and "masterpiece" push harder that way.

Fix. Delete the quality-booster words, name a real capture medium, and ask for the imperfections real images have. On Midjourney, lowering stylisation or using Raw does more than any adjective.

Broken:
Beautiful woman in a cafe, hyperrealistic, 8k, ultra detailed, masterpiece,
cinematic, trending on artstation

Fixed:
Candid photograph of a woman in a cafe, 35mm film, slight grain, mixed available
light, uneven skin texture, one strand of hair out of place, muted colour.

6. "The background contradicts the subject"

What is happening. You described the subject in detail and left the background implied, so the model invented one from the subject's associations. A firefighter gets a fire station whether the brief said a beach or not.

Fix. Give the background its own sentence with its own nouns. If only the backdrop is wrong, do not regenerate: swap it, and our guide to background replacement prompting has the wording per tool.

Broken:
A firefighter in full gear, portrait, beach

Fixed:
Portrait of a firefighter in full turnout gear, waist up, facing camera. Behind
her: an empty grey pebble beach under overcast sky, flat horizon, no buildings,
no fire, no vehicles.

7. "Everything is in focus"

What is happening. Nothing asked for shallow depth of field, so you got the default, which tends toward everything sharp. Focus is a property of the lens, and you did not describe a lens.

Fix. Say the focal length, the aperture and what plane is sharp. Our camera and lens terms reference covers which of these terms genuinely change the render and which are decoration.

Broken:
Close-up of a coffee cup on a table in a busy cafe

Fixed:
Close-up of a coffee cup on a wooden table, shot on an 85mm lens at f/1.8.
Sharp focus on the rim of the cup, the cafe behind dissolved into soft round
bokeh.

8. "My negative words appeared in the image"

What is happening. On a surface with no negative field, a phrase like "no cars" is just a description containing the word cars. Google's own guidance is to use what it calls semantic negative prompts, and its example on the Gemini API image generation page reads, verbatim:

Instead of saying "no cars," describe the intended scene positively: "an empty, deserted street with no signs of traffic."

Midjourney is the exception, with a real --no parameter, though its docs warn that "Midjourney's moderation system reads every word you add to the --no parameter independently" (No parameter, read August 28, 2026), which is how --no modern clothing becomes a moderation flag.

Fix. Say what should be there instead. If your tool has a real negative field, use it there and not in the prompt body. Our negative prompt diagnostic lists which surfaces publish the field at all.

Broken:
A quiet European street, no cars, no people, no signs, no clutter

Fixed:
An empty European street at dawn. Bare cobblestones, closed shutters, a single
lamppost. Nothing parked, nobody walking, no signage on the walls.

Problems 9 to 13: your wording was never the issue

These waste the most time, because the prompt looks wrong and is not. The control lives somewhere else.

9. "The aspect ratio is wrong"

What is happening. On nearly every surface the ratio is a request field, not prose. OpenAI's Images API takes size. Midjourney takes --ar. The Gemini API takes aspect_ratio inside response_format. Typing "16:9" into a sentence sets nothing, and several consumer apps expose no ratio control at all. Note what is documented and what is not: Midjourney says "Some aspect ratios may be slightly changed when upscaling" and that "--ar cannot contain decimals" (Aspect Ratio, read August 28, 2026). No vendor I could find documents silently swapping your ratio for a nearby one.

Fix. Find the field. Our wrong image size diagnostic has the per-surface list.

Broken:
A mountain landscape at sunrise, 16:9 widescreen, wide format please

Fixed:
A mountain landscape at sunrise.
[Midjourney]  ...at sunrise --ar 16:9
[OpenAI]      size: "1536x1024"
[Gemini API]  response_format: { "aspect_ratio": "16:9", "image_size": "2K" }

10. "It ignored my reference image"

What is happening. Two causes. First, references are weighted, and the weight has a default you did not set: Midjourney exposes --iw for image prompts, --cw for character references and --ow for omni references. Second, and more often, the reference is doing exactly what it was designed to do. Midjourney is explicit that it "uses Image Prompts and references as inspiration to guide new creations, not to copy them exactly" (Image Prompts, read August 28, 2026).

Fix. Raise the weight where one exists, and always pair the reference with a text prompt describing the finished scene rather than the edit you want made to the reference.

Broken:
[reference.jpg] make it like this but in a forest

Fixed:
[reference.jpg] A young man with short dark curls and round glasses, standing in
a pine forest at golden hour, three-quarter view, same face and hairstyle as the
reference --cw 100

11. "I get a different image every single time"

What is happening. Randomness is the default, and it is a setting rather than a phrasing. Midjourney publishes --seed and is honest about what it buys you: "Seeds are 99% identical in V8.X" (Seeds, read August 28, 2026). Where no seed is published you cannot pin anything, and OpenAI's image request schemas carry no seed property at all (OpenAI OpenAPI specification, read August 28, 2026).

Fix. Pin the seed where one exists, expecting near-identical rather than identical. Where there is none, stop trying to reproduce and iterate from a saved output as an input image.

Broken:
Same prompt, submitted six times, hoping for the same base composition

Fixed:
[Midjourney]  a lighthouse at dusk, storm clouds --seed 1234 --ar 3:2
[No seed]     re-upload the output you liked as a reference and edit from it

12. "It refused a completely harmless prompt"

What is happening. A moderation classifier, not the image model, and it reads fragments rather than intent. Midjourney's --no example above is the clearest published case of an innocuous phrase splitting into an alarming one.

Fix. Describe what should be present rather than what should be absent, and name garments, ages and actions explicitly so nothing has to be inferred.

Broken:
portrait of a girl, natural look --no modern clothing

Fixed:
Portrait of a young woman in her twenties wearing a linen peasant blouse and a
wool shawl, historical costume, natural light

13. "It falls apart at print size"

What is happening. You are viewing a web-resolution generation at print scale. The detail was never there, and no adjective in a prompt adds pixels.

Fix. Request the highest resolution the surface supports, then upscale deliberately. Words like "8k" in a prompt are style tokens, not resolution controls.

Broken:
Poster artwork of a jazz festival, 8k, ultra high resolution, print quality

Fixed:
Poster artwork of a jazz festival, bold flat shapes, limited palette.
[Set the size or resolution field to the maximum the tool offers,
 then run a dedicated upscale pass on the chosen result.]

Problems 14 to 20: the ones no prompt fixes

Here is the part most guides will not say. These seven are capability limits. You can improve your odds a little, and you cannot cross the ceiling, and the correct move is usually to change tool or add a step rather than to write a better sentence.

14. "The text in the image is gibberish"

Capability limit. OpenAI puts text rendering in its Limitations list: "the model can still struggle with precise text placement and clarity" (read August 28, 2026). Prompting moves this at the margins and no further.

What helps. Keep the string short and quote it exactly. Midjourney requires double quotation marks and notes that "Text generation works best with the standard Latin alphabet, like English letters" (Text Generation, read August 28, 2026). Google's advice is sequencing: "When generating text for an image, Gemini works best if you first generate the text and then ask for an image with the text" (Gemini API image generation, read August 28, 2026).

What actually fixes it. Generate the artwork with blank space where the text goes, then set the type yourself in a layout tool. On which models lead here, see why the text in your AI image is garbled.

Broken:
A vintage coffee shop sign that says Morning Ritual Coffee Roasters, established
2019, single origin beans, hand roasted daily

Fixed:
A vintage hanging shop sign, weathered enamel, reading "MORNING RITUAL" in bold
condensed sans-serif capitals, centred, no other text anywhere in the image

15. "The subject is always dead centre"

Capability limit, partly. Precise placement is the thing OpenAI names explicitly: "the model may have difficulty placing elements precisely in structured or layout-sensitive compositions" (read August 28, 2026). Centre bias is strong, and "on the left" lands somewhere between a suggestion and a coin flip.

What helps. Describe framing as a photographer would rather than as a coordinate.

What actually fixes it. Generate wide and crop, or compose the layout in an editor. If the layout is the deliverable, the image model is the wrong tool for that half of the job.

Broken:
A red canoe positioned in the bottom left corner of the frame with the lake
filling the top right, exactly on the thirds

Fixed:
Wide landscape photograph of a lake at dawn, shot to leave large empty sky.
A small red canoe near the shoreline, tiny in the frame.
[Then crop to the composition you need.]

16. "I asked for five, it drew seven"

Capability limit. Counting is not an operation these models perform. No internal tally is checked against your number, so "five" behaves like a hint rather than a constraint. In my own testing the failure rate climbs sharply past three or four, and identical repeated objects are the worst case.

What helps. Small numbers, and describing an arrangement instead of a count, so the layout carries the number.

What actually fixes it. Generate one object and composite, or accept the count the model gives you and crop.

Broken:
Exactly five red apples on a wooden table

Fixed:
Three red apples on a wooden table, arranged in a row, evenly spaced, each fully
visible and separate from the others

17. "Six fingers again"

Capability limit. Hands are the classic case, and the reason is structural rather than linguistic. "Five fingers" in a prompt is subject to the same non-counting problem as problem 16.

What helps. Simplify the pose. A hand holding a defined object, or resting flat, or out of frame, fails far less often than an open gesturing hand.

What actually fixes it. Inpaint the hand alone rather than rerolling the whole image. Our post on extra limbs and fingers covers why this happens at the training level.

Broken:
A barista gesturing enthusiastically with both hands, five fingers on each hand,
correct anatomy

Fixed:
A barista holding a ceramic cup with both hands wrapped around it, fingers
interlocked at the front, elbows on the counter

18. "Background faces are melted"

Capability limit. Detail is allocated by scale. A face occupying forty pixels does not get a face's worth of structure, which is why crowd scenes look fine until you zoom.

What helps. Keep identifiable people out of the deep background. Ask for backs turned, motion blur, silhouettes or figures cut off by the frame.

What actually fixes it. Bring the face forward so it earns pixels, inpaint each face at scale, or use a real photograph for the crowd.

Broken:
A busy train station concourse with fifty commuters, every face clearly visible
and detailed

Fixed:
A busy train station concourse. In the foreground, one commuter in sharp focus
looking at the departure board. Behind her, blurred figures in motion, faces
turned away or lost in movement.

19. "The character changes between images"

Capability limit. OpenAI states it plainly: "the model may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations" (read August 28, 2026). Describing the person harder does not solve this, because two independent generations share nothing except your words.

What helps. The tool's reference feature, not your adjectives. Midjourney ships Character Reference and Omni Reference with real version restrictions, and warns that "intricate details like specific freckles or logos on clothing might not come out exactly right" (Character Reference, read August 28, 2026). Google advises you to "include previously generated images in subsequent prompts to maintain consistency" (read August 28, 2026).

What actually fixes it. Feed the previous output back in as an input every time, and budget a retouch pass for anything a client sees.

Broken:
Same character as before: a woman, 30s, brown hair, green jacket, freckles,
now in a park

Fixed:
[previous-output.png] The same woman from this image, unchanged face and hair,
now sitting on a park bench in the same green jacket, three-quarter view --cw 100

20. "The edit changed the whole picture"

Capability limit. On prompt-based editing surfaces the mask is guidance, not a hard boundary. OpenAI is unusually direct here: "Masking with GPT Image is entirely prompt-based. The model uses the mask as guidance, but may not follow its exact shape with complete precision" (read August 28, 2026).

What helps. State explicitly what must not change. Every element you want preserved has to be named, because silence is not preservation.

What actually fixes it. A real raster editor with a real selection, for anything where the untouched region must be pixel-identical.

Broken:
Change the sky to sunset

Fixed:
Replace only the sky with a warm sunset, orange to violet gradient. Keep the
building, the trees, the street and the people exactly as they are, including
their existing lighting and colour. Do not reframe or recolour anything else.

What to do when it is a capability limit

Pick one of four exits, and pick it early rather than after twenty rerolls.

A different model. Some are meaningfully better at text than others. None is reliable at it.

An editor. Inpainting a hand, a face or a sign takes two minutes and beats rerolling a whole image.

A real photograph. For a specific person, product or building, a camera is faster and correct.

Typesetting. If the words have to be right, they should not come from the image model. Generate the art, place the type.

None of that is defeatism. It is the difference between an hour of rerolls and a finished asset.

A five-minute diagnostic before you rewrite anything

1. Reproduce it once. Failing the same way twice is systematic. Failing differently each time is randomness, and you need a seed or more samples, not a new sentence.
2. Cut the prompt in half. If the surviving half renders correctly, you had an attention problem, not a wording problem.
3. Check for a field. Ratio, size, resolution, negative prompt, reference weight, seed. If a control exists for your symptom, the prompt was never the lever.
4. Ask the illustrator question from the callout above. Ruler, font file or photograph means capability limit.
5. Only then rewrite, one change at a time, so the result teaches you something.

Most of the value in image prompt troubleshooting is in steps 1 to 4. Step 5 is where everyone starts.

One honest note, since this is our blog. Prompt Architects builds the prompt and never the image. We cannot fix a model's text rendering, and nothing here changes what Midjourney or GPT Image can do. What we do is stop you rewriting the same corrected prompt from memory for the fourth time: Image Prompt Generation starts on Pro at $4.99 a month at the time of writing, per our pricing page, with the Image Prompt Library on Advanced. The Free plan includes 5 prompt enhancements per day, forever, per our FAQ page.