
- Credits per clip
- 300
- Total credits
- 600
- Clips covered by one 600-credit pack
- 2
- 600-credit packs needed
- 1
A failed generation does not consume credits, but a retry is another full-cost job: budget the retries you expect, not the clips you hope for.
Every one of these products fails in the same handful of ways, and almost every failure is an input problem wearing a model’s clothes. The difference between someone who spends 300 credits and gets a usable clip and someone who spends 1,500 and gets nothing is rarely luck; it is which variable they changed between attempts.
Here are the six failure modes, in the order they actually show up, with the input change that addresses each.
1. Identity drift — the face stops being the same face
What it looks like: the first two seconds are right, then the features slide. The nose widens, the eyes migrate, the head settles into a slightly different person.
Cause. The model has a finite amount of information about the face and a growing number of frames to maintain it. Anything that reduces facial information makes it worse:
- a small face inside a wide shot
- motion blur or low light in the source
- a hat, sunglasses, or hair across the face
- a profile or three-quarter angle
- a duration long enough for the drift to accumulate
Fixes, cheapest first:
- Crop the source photo closer, keeping the whole head plus a little shoulder.
- Use a front-facing, evenly lit frame — a phone photo in daylight beats a studio shot with a shadow across one cheek.
- Drop from ten seconds to five. Half the frames, half the drift.
- Keep the camera locked in the prompt. A described camera move invites the model to re-invent the scene, and re-inventing the scene is where the face goes.
- If it still drifts, change the photo, not the prompt.
2. Hand and arm artefacts
What it looks like: fingers fuse, an elbow bends the wrong way, a hand dissolves when it passes the body.
Cause. Fast limb movement plus occlusion. The hand leaves the frame and comes back at an angle the model cannot reconstruct, so it guesses.
Fixes:
- Keep the arms low and the movement small: a sway with bent elbows reads as a dance and stays coherent.
- If the pose matters, use motion control with a reference clip whose arms stay in frame.
- Avoid prompts that add props — a rattle, a balloon, a toy — because the object becomes another thing to keep consistent.
3. Cropped feet, cropped head
What it looks like: the clip is framed so the feet are gone, and with them the joke: a dance without feet is a wobble.
Cause. A landscape source cropped to vertical, or a prompt that does not state framing.
Fixes:
- Say the frame out loud in the prompt: "full body in frame, vertical, 9:16".
- Crop around the subject in the editor before upload, not during generation.
- If the source photo already cuts the feet, choose a composition that reads at the waist — a shoulder shimmy does not need feet.
4. Refused or failed jobs
What it looks like: an error instead of a video.
Cause. Four categories: file validation (size, format, unreadable image), no detectable face, provider-side capacity, and policy. Policy refusals on this product are the usual set — a realistic child subject placed in a sexualised or violent context is refused, as is a photo of a real person used without the right to use it.
What to do: read the message rather than retrying blindly. A validation error is fixed by resizing or converting the image — the image compressor handles that locally and free. A capacity error is fixed by waiting. A policy refusal is not fixed by retrying at all, and retrying it is how accounts get flagged.
Cost note: a refused job does not deduct credits. A completed job that returns a bad result does.
5. Dull movement
What it looks like: the subject sways slightly and does nothing. Technically a dance video; boring.
Cause. Vague prompt language. "Dances happily" carries almost no motion information, so the model picks the safest interpretation, which is the smallest movement.
Fixes:
- Describe the movement mechanically: "two side steps, then a shoulder bounce, repeat".
- Give a count: "on each beat, one bounce".
- Name the intensity relative to something real: "like a slow two-step, not a jump".
- Or stop describing it and use motion control, where the movement comes from a reference clip rather than words.
The difference between a dull clip and a good one is usually one sentence, not a different model.
6. Washed-out or plasticky colour
What it looks like: the result looks like a render, not a photo. Skin is too smooth, the background is unnaturally clean.
Cause. A heavy filter or a heavily edited source photo, plus a model that smooths what it is given. The model is trying to honour an image that has already been processed twice.
Fixes:
- Feed an unedited photo. Not a beauty-filtered one, not a screenshot of a screenshot.
- Keep the background plain but not blank; a wall with a little texture beats a pure white void.
- Do not stack an app filter on top of the generated video.
The retry budget
This is the part that turns a creative problem into an arithmetic problem. At the cheapest setting a clip is 300 credits and the entry pack is 600, so:
| Attempts on the same prompt | Expected result | Credits |
|---|---|---|
| One | One clip, maybe usable | 300 |
| Two, same prompt | Two clips with the same defect class | 600 |
| Two, one variable changed | Two clips, one likely usable | 600 |
| Three, one variable changed each time | High chance of a keeper | 900 |
The planner above lets you price that in advance. The rule that follows from it: change exactly one variable per retry. Photo, then prompt, then duration, then mode. If you change two things and the result improves, you have learned nothing, and the next batch starts from zero again.
A diagnostic order that saves credits
When a clip comes back wrong, work down this list before spending anything:
- Is it drift? → shorten the clip, crop the photo closer.
- Is it a limb? → keep the arms low, remove props.
- Is it framing? → state the frame in the prompt, or crop in the editor.
- Is it dull? → rewrite the movement as mechanics, not mood.
- Is it colour? → feed an unedited photo.
- Was it refused? → read the error; do not retry a policy stop.
Items 3 and 5 cost nothing. Items 1 and 2 cost one generation. That ordering is not an accident: the free fixes come first.
What never to do
- Do not retry the identical prompt. The model is deterministic enough that the same input tends to produce the same defect class.
- Do not generate ten seconds to fix a five-second problem. More frames means more room to drift.
- Do not switch to a bigger mode to fix a prompt problem. Motion-control pro costs 1,000 credits and will faithfully transfer a reference's motion onto a photo that still has a small face in it.
- Do not use a photo you do not have the right to use. No input fix rescues a consent problem — see the consent guide.
Once the input is right, the workflow gets boring in the best way: same photo, same prompt template, same five seconds. Boring inputs are what produce a clip you can publish without paying for a third attempt.
Frequently asked questions
Why does the face change during the video?
Identity drift happens when the model has too little information about the face or too many frames to invent. It is worse with small faces in a wide shot, motion blur, sunglasses or hats that hide features, and longer durations. Fix the input before retrying: crop closer to a sharp, front-facing photo and keep the clip at five seconds.
Why do the hands look wrong?
Fast arm movement is the hardest thing for short-video models to keep coherent, because the hand leaves the frame and re-enters at a different angle. Prompts that keep the arms low and the movement small produce far fewer artefacts; a raised-arm pose is the common cause, not a bad model.
Does a failed generation cost credits?
A job the provider rejects does not deduct credits. A job that completes and returns a bad result does — the model ran, and the cost was real. That is why the cheapest strategy is to fix the input rather than rerun the same prompt, since the same prompt tends to produce the same class of defect.
Why is the movement so small?
Because vague prompts default to a small sway. "Dances energetically" is not a movement description; "two side steps then a shoulder bounce, arms bent at the elbows" is. Motion-control mode is the alternative when you need the movement to come from a reference instead of words.
Why was my job refused?
The usual reasons are a file that is too large, a format the provider does not accept, a photo with no detectable face, or content that the provider's policy blocks — including a realistic child subject in a sexualised or violent context. A refusal is a policy or validation stop, not a model failure, and it is not charged.
![50+ Best AI Baby Dance Video Prompts That Actually Work [2026 Guide]](/imgs/blog/baby-dance-ai-prompts.webp)
![AI Baby Dance Video Generator: How to Make Your Baby Dance with AI [2026]](/imgs/blog/how-to-make-baby-dance-ai-video.webp)







