Prompt troubleshooting: common failures and how to fix them
A symptom-first guide to the problems you will meet with AI images, video and audio, such as warped hands, drifting faces, static clips and wrong text, with specific fixes to try.
Every creator meets the same handful of problems. The picture is almost right but the hands are wrong. The video barely moves. The text on the sign is nonsense. This guide is organised by symptom. Find yours, try the fix, and change only one thing per attempt so you learn what worked.
Before you start, one habit that solves a surprising number of cases: run at the cheapest settings first. A low resolution draft shows you the composition and the motion. Fixing them at 1080p wastes credits.
Images
The hands, fingers or teeth look wrong
These are hard for every image model. Reduce the exposure:
- Compose so hands are partly hidden or small: in pockets, holding a cup from behind, out of frame.
- Avoid prompts that ask for hands doing something precise, such as typing or playing a guitar.
- If one result is close, use Edit Image and describe the fix: "keep the person and background, give her a relaxed hand resting on the table".
The model ignores part of my prompt
Usually the prompt is too long or contradictory.
- Cut it to the five essentials: subject, setting, composition, light, style.
- Put the most important thing first.
- Replace vague words ("nice", "epic", "beautiful") with something visible ("low sun", "rust-coloured cliffs").
- Check for conflicts, such as "minimalist" and "highly detailed intricate" in one sentence.
Counts and positions are wrong
"Three apples" can become five. State the count clearly and keep the scene simple: "exactly three red apples in a row on a white table". If you need precision, generate one and compose in an editor.
Text in the picture is garbled
Keep text short, in quotation marks, and say where it goes: a wooden sign that reads "OPEN". Long sentences, small print and unusual alphabets fail often. For real copy, leave space in the composition and add the text in a design tool.
The same face keeps changing
You are describing a face in words. Use a reference instead. See consistent characters.
The picture looks plastic or over-sharpened
Remove stacked quality words such as "ultra detailed, 8k, masterpiece". Describe a real medium: "35mm photograph, soft natural light, slight film grain". Also check Auto Polish: it appends "highly detailed, cinematic lighting" to your prompt, which can add to the effect. Switch it off if the result is over-processed.
Video
The video barely moves
The prompt probably describes the picture, not the motion.
- Add specific verbs: "turns", "walks", "drifts", "waves roll in".
- Add a camera move such as "slow push-in".
- Add "gently" or "slowly" when you want subtle motion, and leave them out when you want more energy.
- Try a different model. Models differ in how much motion they add by default.
The motion is chaotic or too fast
- Reduce the action to one thing per clip.
- Add "slow", "calm" or "steady".
- Shorten the duration. Drift grows with time.
- Remove stacked camera moves such as "pan while zooming and orbiting".
The subject changes shape or identity
- Start from an approved still with Frame to Video instead of text alone. See the image-to-video workflow.
- Keep the face large, well lit and facing the camera.
- Avoid fast head turns and extreme angles.
Two characters merge or swap
Limit the scene to one main character or give each a clear, different description ("a tall man in a green coat, a small girl in a red hat"). Wide shots with many people are risky.
The audio does not match what I wanted
If the model made its own sound, it may add music or speech you did not ask for. Switch Audio off and add your own with voice-over and music. Or describe the sound in the prompt explicitly.
The aspect ratio is wrong
Set the ratio in Output before you generate. It cannot be changed after. See aspect ratios.
Everything looks soft
Draft resolutions are soft on purpose. Render the final at a higher resolution, and compare, because a higher resolution costs more. A 5 second Seedance 2.0 clip goes from about 444 credits at 480p to 2,363 at 1080p.
Audio
The voice-over sounds rushed or flat
- Shorten the sentences and add full stops.
- Write contractions ("it's", "don't").
- Write numbers and symbols as words.
- Try another voice and model on one sentence before running the whole script.
Words are pronounced wrongly
Respell the word the way it sounds in the script, for example "Siobhan" as "Shiv-awn". Test names and technical terms in a one-sentence run first.
The music does not fit the picture
Describe mood, tempo and instruments, not artists. Turn on Instrumental for anything under speech. Generate the cheap 30 second option first. See the voice-over and music guide.
Generation problems
The button says Get credits
You do not have enough credits for the current settings. Lower the resolution, the duration or the number of outputs, or buy credits. The credits guide explains the price components.
The generation failed
A failed generation is refunded automatically. Try again, or change one setting. Check that your prompt does not ask for something the Terms of Service prohibit, because providers may block those prompts or results. If it keeps failing with different prompts, contact [email protected] with the time and tool.
A tool or model is marked Soon
Some tools and models are listed but disabled, with a Soon label. They cannot be run or charged. Choose a model without the label.
A debugging routine
- Write down what is wrong in one sentence.
- Pick the single prompt change most likely to fix it.
- Run at the lowest resolution and shortest duration.
- If it did not help, undo it and try the next change. Do not stack five changes.
- After two or three failures, change the model or the start image rather than the wording.
Most problems come from too much in one prompt, so the shortest fix is usually to remove something.