Video generation · troubleshooting
Fix the glitches in AI video
Most AI video glitches come from a few causes you can control: characters drift because each clip is generated independently, lengths snap because a model only renders set durations, on screen text garbles because video models draw letters badly, and long videos wander because each clip inherits the last one's errors. The fixes are a repeated character description with reference frames, planning in lengths the model accepts, adding text afterwards, and pinning both ends of every clip.
- 9
- reference frames, at most, to hold a character
- 5 or 10 s
- the only lengths one video model renders
- 15 s
- longest single clip on the realtime model
- 60 s
- videos built by stitching clips
Glitch, cause, fix
Find the row that matches what you see before generating again:
| Glitch | Cause | Fix |
|---|---|---|
| A face changes between shots | Each clip is generated on its own | Reuse one character description word for word, plus reference frames |
| The clip is shorter than asked | The model only renders set lengths | Plan in lengths the model accepts |
| Garbled words on screen | Video models draw letters badly | Add titles and captions as overlays afterwards |
| A long video slowly wanders | Each clip starts from the last frame and inherits its drift | Pin both the start and the end of every clip |
| A soft, smeared face | The face is small in frame | Crop closer; resolution will not fix it |
| An unexpected cut mid clip | The model composed several shots | Describe one continuous shot |
Character drift, and why prompts should repeat themselves
Nothing carries between clips on its own, so a person can quietly become someone else by the third shot. Improvising a fresh description for every clip is what causes it. Register each recurring character once, with one canonical description, and paste it almost word for word into every prompt that character appears in.
Reference images do the rest. When Percify remakes a short, up to nine keyframes from the source are passed to the model so characters stay the same people across clips. For your own characters, start every clip from the same saved avatar image.
Lengths are set by the model, not by you
Video models accept a fixed menu of durations. One renders exactly 5 or 10 seconds and nothing else; the realtime model renders 5, 10 or 15. Ask a model for a length it does not offer and it will render the nearest one it does, so a plan built on nine second clips quietly delivers five. Percify refuses a total it cannot render exactly rather than delivering a shorter video than you paid for.
Long videos: why they drift, and how pinning fixes it
A long video is several clips stitched together. The simple way starts each clip on the last frame of the previous one, so any drift in one clip is passed to the next and compounds. The robust way plans the story as a series of stills, each an edit of the one before, and generates every clip between two pinned stills. Nothing is taken from the rendered video, so drift cannot pass forward. Both are behind long AI video.
Text, refusals and failed runs
Put words on screen after generation, as overlays. Generated lettering renders badly and dates instantly. If a prompt is refused by content moderation, see why AI video prompts get rejected. When a generation fails outright, the credits are refunded automatically.
Clean up a glitchy video in four steps
- 1
Name the glitch
Match it to a row above; each has a different cause.
- 2
Fix the input
One saved character image and one canonical description, reused every time.
- 3
Plan lengths the model accepts
Build long videos from clips of the lengths the model renders.
- 4
Add text last
Titles and captions go on as overlays after the video is generated.
One character, many shots
Real Percify output: the same mascot shown from several angles.
Multi angle scene
one mascot, several views
Multi angle scene
the same character, new shots
Tyler
a 3D character
Frequently asked questions
Why does my AI character change between clips?
Each clip is generated independently, so nothing carries over unless you carry it. Reuse one canonical description word for word and start every clip from the same reference image.
Why is my AI video shorter than I asked for?
Video models only render certain lengths, such as exactly 5 or 10 seconds. A request for another length is rounded to one the model offers, so plan in the lengths it accepts.
Why is text garbled in AI video?
Video models draw letters badly. Add titles, captions and labels as overlays after the video is generated.
Why do long AI videos drift?
When each clip starts from the last frame of the one before, errors pass forward and compound. Pinning both ends of every clip to planned stills stops the drift from carrying.
Do I get credits back when a generation fails?
Yes. On Percify, a generation that fails is refunded automatically.
Keep reading
Fix the input, not the retry
Repeat the description, plan the lengths, add the text last.