Skip to main content
seedanceseedance 2.5troubleshootingprompt not workingai videodebugginggeneration failed

Seedance Ignoring Your Prompt? The Real Causes (and Fixes)

A fault finder for AI video generation: which symptom means a wiring bug, which means a content refusal wearing an outage costume, which means your prompt, and which means the model simply will not do it.

Brian Bautista · Co-Founder & Creative Director|August 17, 202610 min read

Quick answer

Diagnose before you rewrite. An output that ignores your reference completely is almost always a wiring problem, not a prompt problem. Repeated identical failures with a generic internal error are often a content refusal, so retrying wastes every attempt. Blurry or warped motion usually traces to the word fast. A rushed middle means no beat structure. And a subject that will not hold still is a property of generative video that no prompt fixes.

Most troubleshooting advice for AI video amounts to "change one thing and try again." That is correct and useless on its own, because the hard part is knowing which thing.

The useful move is to sort the failure into a class first. A wiring bug, a content refusal, a prompt conflict, a bad source photo and a hard model limit all look similar from the outside (you asked for something and did not get it) and they need completely different responses. Retrying is right for exactly one of them.

This guide is the debugging half of the Seedance prompt handbook. It comes out of running a production pipeline with an automated failure classifier, thousands of paid generations, and the refunds that taught us to tell these classes apart. That pipeline runs Seedance 2.0, not 2.5, and provider behaviour varies, so treat the provider-specific items as patterns to check rather than universal truths.

The five classes

ClassIt looks likeRetrying helps?
PlumbingThe output ignored an input entirelyNo. Fix the wiring.
ModerationRepeated identical failures, generic errorNo. It will refuse every time.
PromptMost of it worked, one instruction lostOnly with a change.
Input photoLikeness drifts, subject will not settleNo. Change the photo.
Model limitIt has never worked, for anyoneNo. Change the approach.

The single most expensive mistake is treating a moderation refusal or a plumbing bug as a transient error and retrying. Those attempts are guaranteed to fail, and on a paid API they are guaranteed to cost.

The fault finder

Symptom to cause to one fix15 symptoms
The output ignored my reference photo completely. Different subject, different setting, like it never saw it.Plumbing

What's actually happening

Almost never the prompt. Either the reference parameter name was wrong, or the reference was uploaded but never named in the prompt text. Some APIs silently ignore unknown input keys rather than erroring, which quietly downgrades an image-to-video call into text-to-video. The result looks exactly like a prompt-adherence failure, which is why people rewrite the prompt for an hour.

Change this one thing

Check the wiring before touching the words. Confirm the parameter name against the provider's current docs, and confirm the prompt names the reference with the right tag and capitalisation.
The wrong reference got used. It applied the location as the character, or the style image as the subject.Plumbing

What's actually happening

References are numbered by upload order, not by type. Everyone assumes images and videos have separate counters. They do not, so an upload order you did not intend silently renumbers every tag in your prompt.

Change this one thing

Reorder the uploads so the reference you most need preserved is first, then re-check every tag number in the prompt against the new order.
It keeps failing with a generic Internal Error, please try again later.Moderation

What's actually happening

That string is sometimes a content refusal wearing an outage costume. Some providers wrap an upstream model's moderation response in a generic error, so it reads as transient and every retry goes straight back to the model that just refused. Photos of children and certain likeness edits are common triggers.

Change this one thing

If three attempts fail identically and quickly while other jobs succeed on the same provider, treat it as a refusal. Change the input photo or route the job to a different model. Do not keep retrying.
One specific photo fails every time. Everything else works.Moderation

What's actually happening

That is the signature of a refusal rather than an outage. A genuine outage fails everything; a refusal fails only the job carrying the input that triggered it. Different models draw these lines in different places, and a photo one model refuses another will often render without comment.

Change this one thing

Swap the photo, or swap the model. Both are faster than arguing with a classifier you cannot see.
Motion looks blurry, smeared or warped whenever anything moves quickly.Prompt

What's actually happening

Usually the word fast. It is the single keyword most likely to degrade motion quality, because it asks for speed without saying what moves or how, and the model resolves that ambiguity with blur.

Change this one thing

Delete the word fast and describe the motion concretely instead. Whip pan, explosive burst, snaps into position, breaks into a sprint. Same energy, far fewer artifacts.
The clip rushes through the middle and lingers on the last shot.Prompt

What's actually happening

Nothing told it how to spend the time. A prompt written as one continuous description leaves duration allocation to the model, and on longer clips it compresses the middle.

Change this one thing

Split the clip into three to five timed beats with one job each. On a 30-second generation this single change fixes more than any other.
Locations morph into each other instead of cutting.Prompt

What's actually happening

No transition was stated, so the model interpolated between two descriptions rather than cutting between two shots. Smearing between locations is the most recognisable AI-video artifact there is.

Change this one thing

Name the cut explicitly. Two words at the head of the beat, like Inside the car, is usually enough to turn a morph into an edit.
The camera ignored my direction completely.Prompt

What's actually happening

Almost always competition. Two camera instructions in one beat, or a camera instruction fighting a subject-movement instruction, and the model averages them into neither. Vague direction has the same effect: dynamic camerawork carries no information.

Change this one thing

One camera instruction per beat, stated concretely, and say where the subject sits in frame. Keep the runner in the left third of frame is followed; dynamic tracking is not.
The output drifted toward illustration or animation when I wanted photoreal.Prompt

What's actually happening

Common when the reference image is itself generated or stylised: the model takes the rendering style from the reference as well as the content.

Change this one thing

Add 真人实拍 (real-person live-action footage) near the end of the prompt. Near the end matters, see the next row.
The video came back with Chinese dialogue I never asked for.Prompt

What's actually happening

真人实拍 placed at or near the start of the prompt. Early placement makes the model read the whole scene as Chinese-language content, and it generates dialogue to match.

Change this one thing

Move it to the end of the prompt, just before the constraints block. Same photoreal push, no unexpected language.
My long negative prompt is not working, or made things worse.Prompt

What's actually happening

Naming a thing puts it in context whether or not you asked for it, and a long generic avoid list dilutes the instructions that matter. A twelve-item list of things you do not want competes with the six things you do.

Change this one thing

Cut the avoid list to the artifacts that actually show up in your outputs, and spend the words on what you do want instead: which reference controls what, and what stays unchanged.
The face drifts even though I attached a good reference.Input photo

What's actually happening

If your workflow builds an intermediate image first, that step re-rendered the person rather than compositing them, so the likeness shifted before the video model saw anything. The video is faithfully rendering an intermediate that already drifted.

Change this one thing

Attach the original photograph as an additional reference alongside the generated one, or skip the intermediate entirely where the shot allows it.
My subject will not hold still, no matter how forcefully I write frozen or does not move.Input photo

What's actually happening

This is a property of generative video, not a prompt bug. We proved it across four renders and roughly $4.86, restating the freeze instruction five times in one attempt, then reproduced the identical failure on an unrelated model family. What actually varied was the source photo: every failure started from someone mid-stride, and the model was completing the step the frame implied.

Change this one thing

Stop prompting for it. Start from a photo of a held pose or gesture, which has no motion to finish. For a genuinely perfect freeze, use a depth-parallax renderer, where nothing is generated so nothing can drift.
Text, logos and signage come back garbled.Model limit

What's actually happening

Small text is one of the weakest areas of every current video model, and text baked into a reference image tends to survive into the video as approximate lettering.

Change this one thing

Do not ask the model for text you need to be correct. Leave the surface blank and composite the text in post, where it will be sharp and spelled right.
Audio is generic ambience even though I enabled it.Model limit

What's actually happening

Enabling audio is not directing it. generate_audio defaults to on, and a prompt that says nothing about sound gets a generic bed.

Change this one thing

Describe the sound where it happens: rain on the roof, wipers squeaking, muffled radio. Write dialogue with its delivery. If you attached a music reference, say what syncs to it.

The discipline that makes debugging cheap

Change one variable at a time, and fix the seed while you do it. If a generation is 90% right and you rewrite the whole prompt, you have introduced five new variables to fix one problem. Find the failure, change the single instruction connected to it, regenerate. Controlled iteration teaches you how the model behaves; wholesale rewriting teaches you nothing.

Draft cheap. Iterate at the lowest resolution the provider offers and finalise higher. Billing is per second, and on Seedance 2.5 it also scales with reference count, so a heavy reference stack makes every draft more expensive. That is one more argument for nine references rather than forty.

Keep a second model on the same provider. When something fails in a way you cannot explain, throwing the identical prompt at a different model is the fastest way to separate a model problem from a provider problem. Kie.ai carries Seedance and MiniMax H3 through one API, which makes that comparison a one-line change rather than a new integration.

Things Seedance will not do

Honest limits, so you stop spending money proving them:

  • Hold a subject completely frozen while the camera moves. Covered above. These models animate people. Prompting harder does not help; a different rendering technique does.
  • Produce reliably correct small text. Signage, scoreboards, UI labels, anything you need spelled right. Composite it afterwards.
  • Native 4K. Despite a lot of pages saying otherwise, Seedance 2.5 generates at 480p and 720p with 1080p arriving around now. The 4K on offer is an upscale, which is a different thing and looks like one. If native resolution is your constraint this month, MiniMax H3 does native 2K.
  • Hold more than about eight identities in one clip. Past that, restage the shot rather than adding references.
  • Read your mind about which reference does what. It will not reliably infer that image 3 is the location. Say so, every time.

Where to go next

We built an automated failure classifier for Starrd because sorting these classes by hand did not scale past a few hundred generations a week. Most of what is above is that classifier's taxonomy, written out in English.

Frequently Asked Questions

Why is Seedance ignoring my prompt?

Work out which layer failed before rewriting anything. If the output ignored a reference image entirely, that is usually a wiring problem: the reference parameter was wrong, or the reference was never named in the prompt text. If the model followed most of the prompt but missed one instruction, that is usually competition, where two instructions in your prompt contradict each other. If it failed to generate at all with a generic error, it may be a content refusal rather than an outage.

What does Internal Error, please try again later actually mean?

Sometimes exactly what it says, and sometimes a content refusal in disguise. Some providers wrap an upstream model's moderation refusal in a generic error string, which reads as transient. The tell is that every retry fails identically and quickly while other jobs on the same provider succeed. A genuine outage fails everything; a refusal fails only the job with the specific input that triggered it.

Why is my AI video blurry or warped during fast movement?

Most often the word fast itself. It is the single keyword most likely to degrade motion quality across Seedance generations, because it asks for speed without describing what is moving or how. Replace it with concrete motion language: whip pan, explosive burst, snaps into position, breaks into a sprint. You get the same energy with far fewer artifacts.

Why does my video rush through the middle and linger at the end?

Because nothing told it how to spend the time. A prompt written as one continuous description leaves the model to allocate duration itself, and on longer clips it tends to compress the middle. Split the clip into three to five timed beats with one job each and the pacing problem disappears.

Do negative prompts work in Seedance?

Short, targeted ones do. Long generic ones dilute and can backfire, because naming a thing puts it in the context whether or not you asked for it. Keep an avoid list to the artifacts that actually appear in your outputs, and prefer stating what you want. Instead of a long list of things not to do, say which reference controls what and what stays unchanged.

Why won't my subject hold still even though I wrote frozen?

Because these models animate people; that is what they are trained to do. We tested this exhaustively, four renders and about $4.86, restating the freeze instruction up to five times in one prompt, and then reproduced the identical failure on an unrelated model family. The variable that mattered was the source photo, not the prompt: a photo of a held pose holds, a photo of someone mid-stride keeps moving because the model completes the implied action.

About the author

Brian Bautista · Co-Founder & Creative Director

Brian is co-founder and creative director at Starrd, working as a creative technologist and data scientist. He tracks viral AI-video trends, designs Starrd's scene templates, and writes the deep-dive model comparisons and prompting breakdowns.

Related Articles

Ready to create your own video?

Pick a template, upload your photos, and generate a cinematic AI video in minutes.

Browse Templates