Why AI Generations Fail, and How to Fix Each Problem
Most disappointing AI results are not random. A warped label, a sixth finger, a word spelled wrong or a bottle that melts halfway through a video each point to a specific cause, and usually to one change that fixes it.
The short answer: name the defect, change only the thing that causes it, and keep everything that worked. Do not throw the whole prompt away. If the same fix fails twice, switch to a model that is better at that job. And there are two very different kinds of failure: a result you do not like (a prompt problem) and a generation that actually failed (a job that errored or never finished). The second kind gets its credits back automatically once our systems confirm it failed.
Below is a quick-fix table, then each problem in more detail, then how to fix a result with [EB-1](/agents/eb-1) in one message.
The quick-fix table
Find your problem in the left column and make the change on the right. Change one thing at a time so you know what worked.
| What you see | Likely cause | What to change |
|---|---|---|
| The product changed shape or the label is wrong | The product was described, not shown, or the scene is crowded | Attach the product photo (video: as the start frame), say "identical shape, label and colours", remove competing objects |
| Garbled or misspelled text | A photo model was asked to write words | Use GPT Image 2, quote every word exactly, keep the copy short |
| Extra or warped fingers and hands | Too many hands, vague grip, complex pose | Fewer hands, describe the exact grip, a simpler pose, a closer crop |
| The face changes between ads | No reference image, wording changed | Reuse the same character photo and copy the description word for word |
| The video morphs, melts or flickers | Too much motion for the length, busy patterns | One slow camera move, fewer subjects, longer beats, a start frame |
| The mouth is out of sync, speech is rushed | Too many words for the seconds | Fewer words (about 2.5 per second), one speaker, face to camera |
| It looks fake, cheap or "AI" | Generic adjectives, flat lighting | Real materials, one motivated light, a real lens, natural skin texture |
| It missed the brief | The key subject or action came late in the prompt | Put the subject and action first, with exact counts |
| Important parts hidden on TikTok or Reels | Content near the top or bottom of the vertical frame | Keep the subject centred, nothing key in the top 15% or bottom 25% |
Is it a bad result or a failed generation?
They look similar from your chair but are handled differently.
- A failed generation errored at the provider, never returned anything, or got stuck. When our systems confirm the failure, the credits for it are returned automatically. You do not need to ask.
- A result you do not like is a finished generation that did not match what you wanted. Once a provider has accepted a job and delivered it, credits are not returned for taste, which is why it pays to fix the prompt carefully before running it again.
If a generation seems stuck for much longer than usual, give it a little time, then check your [Assets](/assets). If it still has not appeared, contact [support](/support) with the time and the tool you used.
Why did my product change?
Models are trained on millions of products, so a vague description gets you a plausible product, not yours. The fix is to show it:
- Images: attach the product photo as a reference. When there are several photos, say which one is the product.
- Videos: use the product photo as the start frame, so the clip opens on your exact product, and describe only what moves.
- Say what must stay the same: "identical shape, label, colours and proportions."
- Remove objects that compete with the product, and keep the label facing the camera in key shots.
On the video models Grok Imagine and Kling 4.0 Flash, reference photos are not used: only a start photo is. Pass the product as the start photo, or switch to Seedance 2.5.
Why is the text garbled?
Photo models are built to make pictures, not to typeset. For anything with readable words, use GPT Image 2 (in EB-1, the "text" quality), and write the prompt like a layout spec:
- every word in double quotes, exactly once;
- where it sits and how it looks: "headline across the top third, bold condensed sans, white";
- one headline under eight words, at most one sub-line and one call to action;
- "all text spelled exactly as quoted, nothing else written."
For video, leave text out of the generation and add captions in your editor.
Why are the hands wrong?
Hands go wrong when a prompt leaves too much open. "Holding the product" could be one hand or two, from any side, at any angle. Be specific and simple:
- say how many hands: "one right hand";
- say the grip: "holding the bottle from below, thumb on the front";
- say what else they do: nothing, ideally;
- crop closer, so the hand is large and clear, or crop the hands out if they are not the point.
Why does my video morph or flicker?
The model is trying to do too much in the time it has. Fast spins, many subjects, busy patterns and several actions in one beat all push it to cut corners.
- Use one camera move per shot, and make it slow: "slowly pushes in".
- Fewer subjects. One person and one product is easiest.
- Give each beat at least three to five seconds. If the story is rushed, choose a longer length or cut a beat.
- Start from your product photo so the first frame is solid.
- Avoid fine stripes and dense patterns, and keep light steady.
Why is the voice out of sync?
Lip-sync suffers when there are too many words for the seconds available. Budget about two and a half words per second, and keep the first beat and the end card free of talking. A 10-second clip holds about 20 words; a 15-second clip about 30.
Also: one speaker at a time, facing the camera, in a medium close-up. Write the exact line in quotes, with who says it and how (age, tone, accent). For talking clips, use a Seedance model; Seedance 2.5 is the strongest for natural speech. Some other models may not generate sound at all, so do not build a talking ad on them.
Why does it look fake?
"Perfect", "flawless" and "stunning" push a model toward a glossy, plastic look. Real photographs have specific materials, light from a direction you could point to, and small imperfections.
Replace adjectives with things a camera would see: brushed aluminium, linen, a slightly messy real kitchen, 8 a.m. window light from the left, an 85mm lens, natural skin texture, a little film grain. One motivated light source almost always looks more real than "studio lighting".
How do I fix a result with EB-1?
In [EB-1](/agents/eb-1), select the result on your board and say what is wrong in plain words: "the label is warped and she has six fingers." EB-1 reads that as an edit of that result, keeps what works, names the defect and changes only what causes it. It then proposes the fix with its price, and nothing runs until you approve it.
EB-1 also reviews finished images against your brief before you see them, and shows a quality note on the card, so you can fix a weak result in one tap.
What matters most
- Name the defect before changing anything.
- Change one thing, keep everything that worked.
- Show the product, do not describe it.
- Words belong to GPT Image 2 or your editor.
- Simple, slow, specific beats beat busy ones.
- If the same fix fails twice, change the model.
Questions people also ask
Do I get my credits back if a generation fails?
Yes, when our systems confirm the generation failed, its credits are returned automatically. Credits are not returned for finished results you simply do not like.
My generation has been running for a long time. Is it stuck?
Some videos take several minutes. Check your Assets after a while. If nothing has appeared much later than usual, contact support with the time and tool so we can check the job.
Why did my request get blocked?
Ecombos screens requests and results for sexual content, minors, real people without permission, hate and deceptive content. Blocked requests are not charged. Rephrase the request if it was blocked by mistake, or contact support.
Should I start over when a result is bad?
Usually not. Keep the prompt, change the one thing causing the problem, and run it again. Starting over throws away everything that worked.
Which model is best for text in images?
GPT Image 2. Quote the exact words and where they sit.
Which model is best for talking videos?
A Seedance model. Seedance 2.5 handles natural speech and lip-sync best, and runs up to 30 seconds.
Can EB-1 fix results it did not make?
You can attach any of your own images with @ and ask for an edit. EB-1 will propose the change and its price before anything runs.