Troubleshooting
Common AI video failures, why they happen, and the fix to try before you spend another credit.
Most failures come from asking one shot to do too much. Before you regenerate, find the problem in the table below and change the prompt or the shot, not just the seed. If the same failure shows up three times, rewrite the shot (see Generate and iterate).
Common failures
| Problem | Why it happens | Fix |
|---|---|---|
| Faces morph or drift | Nothing anchors identity; long clips; fast head turns; people crossing in front | Attach a character sheet or reference to every shot; start from a clean first frame (image to video); keep shots short; one action at a time; on Seedance, reference strength around 70 to 80% (single source) |
| Hands look wrong | Hands are small, fast and complex | Frame hands out or keep them still ("hands resting on the table"); slow the motion when hands must show; use a reference clip for a specific gesture; fix the hand region in post |
| Physics breaks (objects pass through each other, liquids act strangely) | The model predicts frames, it does not simulate; errors build up over time | Describe the result of the action ("debris scatters on impact"); one action per clip; generate 3 to 5 second pieces and join them |
| Text and logos are garbled | Video models render text poorly | Add text in the edit; attach logos as images, never describe them in words; on Seedance keep text short, in one language, and state its position and timing |
| Quality drifts in long clips | Identity, lighting and motion drift past about 8 to 10 seconds | Keep shots short; for longer Seedance shots, write a per second timeline; extensions are not frame exact, so plan a pause or cut at the join |
| Music or sound you did not ask for | An empty audio slot gets filled; Seedance 2.5 ignores the audio reference or invents sound in roughly 15 to 20% of runs (one creator) | Describe the audio and write "no music"; ask for sound effects only and add music in the edit; replace the audio in post with a tool like Mirelo |
| Prompt rejected by moderation | Real faces, blood, words like "shirtless"; Flow and Seedance are strict | Use generated or stylized characters (Seedance 2.5 does not accept real faces in references); describe the clothing that is there; try Kling for action that Seedance rejects |
| Character changes between shots | The model has no memory between generations; references conflict; the style line is paraphrased | Reuse the same character sheet and the same style line word for word in every shot; settle one costume design; restate the key traits in each prompt; on Kling, use Elements built from 2 to 4 images |
| Camera ignores your instructions | Several moves stacked in one prompt; no lens named; camera described by position | One camera move per shot; say "static, locked-off" when it should not move; name a lens; describe what is visible in frame rather than where the camera stands; block the camera in 3D first for exact moves |
| Part of the prompt is ignored | Prompt too long; later instructions dropped | Put subject and action first; cut to 60 to 100 words; give a stubborn detail its own sentence or a reference image |
| Jitter or stutter | Too many or vague moves | Use pace words (slow, steady); one move; say what stays still |
| Lip sync drifts or voices merge | Long lines; unlabeled speakers; weak source face | Short lines split across cuts; label each speaker with a tone; supply your own voice track; frontal, high resolution face |
| Kling stuck at 99% | The model does not know how to end the motion | Add an end state ("then settles"); 5 seconds instead of 10; lower resolution; try off peak hours (single source) |
| First frames look soft | Common in generations | Trim the first few frames in the edit |
| Storyboard numbers appear in the video | Labels were drawn inside the panels | Keep labels outside the panel art; add "no text, numbers, captions" to the video prompt |
| Stylized character turns photoreal (Veo) | Veo leans toward realistic texture | State the style limits: flat shading, ink outlines, no photoreal texture |
Sources rated "consensus" in our research: faces, hands, physics, text, unwanted audio, ignored prompts and jitter. The rest come from one to three sources.
Moderation rejections cost nothing on some platforms
Seedance (BytePlus API), Kling (API), MiniMax and Vidu do not charge for failed or rejected generations (checked 15 Oct 2026). Users still report losing hours to repeated flags on Flow and Seedance, so rewrite the prompt instead of resubmitting it unchanged.
Fix prompts
Copy these and replace the names and details with yours.
For a face that keeps changing:
MAYA from @image1, face fully visible, facing camera at 3/4,
minimal head movement, slow and steady motion, no other people crossing
in front, static camera, 4 seconds.For hands that must be on screen:
She rests both hands flat on the table, fingers still, and only her eyes and
head move as she looks up; slow motion; hands stay visible and relaxed.For motion that will not stop, or a Kling job stuck at 99%:
...the flag ripples twice in the gust, then the wind dies and it hangs still
for the final second; camera holds.For unwanted music or voices:
Audio: rain on window, low fridge hum, single cup set down at 3s. No music,
no voices.Fix it in post instead
Some problems are cheaper to fix in the edit than to regenerate.
- Small error in a good take: try an edit model (Gemini Omni Flash, Runway Aleph) with a short prompt that ends with "Keep everything else the same".
- Flicker on medium paced shots: gentle frame interpolation. Avoid it on fast motion, where it ghosts. See Finishing.
- Clips that do not match in color: match to one hero clip, then apply one LUT across everything. See Editing.
- A bad moment in the middle of a shot: cut away to a reaction or detail shot and back.
- Wrong or invented audio: mute it and rebuild the sound with separate layers. See Audio.
When to stop
If a shot needs several characters interacting, detailed hand work and readable text all at once, it may not be possible with today's models. The Big Yowie creator reports multi character scenes taking more than 50 prompts. Split the idea into simpler shots.
Sources
- Morphic, AI video troubleshooting
- Hackernoon, fixing warped hands and faces
- heyuan110, Seedance 2 prompt guide
- fal, Kling 3.0 prompting guide and Seedance 2.0 prompting guide
- VEED, Kling prompting guide
- elser.ai, Kling vs Seedance vs Veo for anime
- Kapwing, storyboard to video with Seedance 2.5
- invideo, why clips are limited to 8 to 10 seconds
- BytePlus ModelArk Seedance 2.5 docs
- OriSilver on X, Seedance audio
- JXP, GPT Image 2.5 storyboard test