AI video models are sensitive to how you write your prompt. If your generations aren't matching what you asked for — sped-up audio, dialogue that doesn't match your script, characters doing the wrong thing — the prompt is usually the cause rather than the model. This guide covers what to change.
Turn off Enhance for dialogue
The Enhance toggle sits next to the generate button and is on by default. When it's on, VEED rewrites your prompt before sending it to the AI model, and the model sees the rewritten version rather than yours.
That's helpful for short, loose prompts. It's a problem when your prompt contains exact dialogue, because the rewrite can change or drop your lines.
If your prompt contains dialogue you want spoken word for word, switch Enhance off.
One sign this is happening: the file name of your downloaded video doesn't match anything you typed. Download file names come from the prompt the model actually received, so if yours reads like an instruction you never wrote, Enhance rewrote it.
Match your dialogue to your clip length
Speech runs at roughly 2.5 words per second. If you write more dialogue than fits the clip, the model has no choice but to speed the audio up to fit, which is why generations come back sounding rushed or garbled.
As a rough guide:
5 seconds — one short line, around 12 words
10 seconds — around 25 words
15 seconds — around 35 words
One or two lines per clip works best. For anything longer, generate several clips and join them in the editor.
Keep each scene focused
Describe one main action or sequence at a time. A prompt combining several characters, multiple actions, camera movements, and a long script gives the model too much to resolve at once, and something usually gets dropped.
If your scene has all of those, split it into separate clips and assemble them in the editor. Each clip gets the model's full attention, and you only regenerate the part that didn't work.
Write only what should be seen and heard
The model treats everything in your prompt as content to render. Stage directions and instructions aimed at the AI get interpreted as part of the scene.
Leave out:
Stage directions such as (quietly) or (to Jerome)
Instructions to the model such as "do not adlib" or "these characters have a strict script"
Anything conflicting or repeated
Include: who is in shot and where, the mood, any camera movement, and the exact lines you want spoken.
A prompt format that works
One sentence describing the scene, then each line of dialogue attributed by name:
Two people seated at a conference table, Dr. Mila on the right, Raymond on the left. Solemn, restrained mood.
Dr. Mila says: "Their leaving was a signal the profession refused to read."
Raymond replies: "I'm losing them in my memory too."
If you're working from a reference image, keep using it — it helps the model hold your characters consistent between clips.
Choose the right model
Models differ in what they handle well, so it's worth switching if one isn't working for your scene.
Seedance supports longer generations — up to 15 seconds in AI Playground — which suits scenes with more dialogue or action.
Veo applies stricter content filtering, so prompts that generate fine elsewhere may be blocked.
If you're generating multi-character dialogue, Seedance 2 at 15 seconds is usually the best starting point. Cost varies by model and is shown before you generate.
What to expect
Lip-synced dialogue with several characters is the hardest thing to generate. Even a well-written prompt often needs two or three attempts per clip. That's the current state of the models rather than a fault in your prompt.
Successful generations use credits, even if you don't like the result. Credits are spent when a generation completes, not when it meets your expectations — which is why it's worth getting the prompt right before running it repeatedly.
Some prompts are blocked by the model provider, not by VEED. These safety filters sit with the companies that build the models, and we can't override them. Rephrasing, or trying a less restrictive model, will often get you through.
FAQs
Q: Why is the audio in my video sped up?
There's more dialogue in your prompt than fits the clip length, so the model compresses the speech. Shorten the lines, or generate a longer clip and split the scene across several.
Q: The video doesn't use the dialogue I wrote. Why?
Most likely Enhance was on and rewrote your prompt before the model saw it. Switch it off and generate again.
Q: Why does my downloaded file have a name I didn't write?
File names come from the prompt the model received. If it doesn't match what you typed, your prompt was rewritten by Enhance.
Q: Can I get my credits back for a generation I'm not happy with?
No. Credits are used whenever a generation completes successfully, regardless of the result. Only failed generations cost nothing.
Q: My prompt was blocked. What can I do?
Content filtering is applied by the model provider rather than VEED. Try rephrasing the parts likely to have triggered it, or switch to a different model. For example, Veo filters more strictly than Seedance.
Q: How do I join several clips into one video?
Generate each clip separately, then add them to a project in the editor and arrange them on the timeline. See how to cut and trim your videos.