The most common AI video prompt mistakes are describing a still photo instead of describing motion, stacking three or four actions into a five-second clip, dropping identity-preservation language on image-to-video, burying negatives inside the positive prompt, and re-rolling the same prompt instead of changing one variable at a time. All are cheap to fix, and on VdoBloom all of them cost you credits, because credits are charged on submission and not on whether you liked the result.

Below are the mistakes we see most often across the 105+ models on the platform, with the exact rewrite for each. The fixes are drawn from how our own shipped effect templates are written — you can inspect those yourself by opening any effect tab and reading the prompt that loads into the box.

The mistakes, at a glance

Mistake	What you actually get	The fix
Describing the photo, not the motion	A near-frozen clip with a slow zoom	Name one verb and one camera move
Three actions in a 5-second clip	Rushed, warped motion; limbs smear	One action per clip, or buy 10s
No identity lock on image-to-video	The face drifts by second three	Add the preserve-features sentence
Negatives in the positive prompt	You summon the thing you banned	Use the negative prompt field
Prompt written for the wrong aspect	Subject cropped at the knees	Set the aspect ratio control, not the prompt
Overwriting an effect template	Loses the tuning that made it work	Edit one clause, keep the scaffold
Iterating on the 1080p tier	203 credits per failed attempt	Test at 480p, finalise at 1080p
Re-rolling an identical prompt	Same failure, minus the credits	Change exactly one variable

Mistake 1: you described the photograph, not the movement

The biggest one. A prompt like “a woman in a red dress standing on a balcony at sunset, cinematic, 8k, highly detailed” is an image prompt — a frozen instant. Video models read it, find no verb, and give you the safest interpretation: a barely-moving frame with a slow push-in.

The fix: every video prompt needs at least one subject verb and one camera instruction. Rewrite it as “a woman in a red dress turns from the balcony railing toward camera, her hair lifting in the wind; a slow push-in settles on her face as the sun drops.” Same scene, but now the model has something to animate.

Mistake 2: you asked for four things in five seconds

“She walks in, sets down her bag, pours a coffee, then looks out the window” is roughly twenty seconds of screen time. Ask a 5-second model for it and it compresses — that is where the rubber-limb, sped-up look comes from.

The fix: one beat per clip. If you genuinely need the sequence, generate it as separate clips and cut them together, or step up to a 10-second tier. On the WAN 2.7 text-to-video ladder, 720p 5s costs 32 credits and 720p 10s costs 64 — you are paying exactly double for exactly double the runway, so there is no penalty for buying the length you actually need.

Mistake 3: you forgot to lock identity on image-to-video

Upload a photo, write “she dances,” and you will often get someone who is recognisably not the person you uploaded by the end of the clip. The model is free to reinterpret anything you did not pin down.

The fix: use the sentence pattern our own effect templates use. Every shipped template on VdoBloom contains a variant of: preserve facial features, body proportions, clothing, lighting, and identity exactly as in the original image. Several also add: only the people from the uploaded images may appear in the scene — no additional persons or characters. That second clause is what stops phantom extras walking into frame. Copy both into your custom prompts. If the photo is of a real person other than yourself, you must have that person’s consent before animating them.

Mistake 4: your negatives are in the positive prompt

“No text, no watermark, no extra fingers” inside the main prompt tells the model those tokens are relevant to the scene. Video models are bad at processing negation, so you frequently get the exact artefact you banned.

The fix: the video creator has a separate negative prompt field. Put exclusions there and keep the main prompt purely affirmative.

Mistake 5: you tried to fix framing with words

“Full body shot, vertical, 9:16” in the prompt text does far less than the aspect ratio control. Framing, resolution and duration are structured settings sent to the provider as parameters; the prompt cannot override them.

The fix: set aspect ratio, resolution and duration in the controls, and check the model’s supported list on its page in the model directory first — not every model supports every ratio.

Mistake 6: you deleted the template instead of editing it

Effect templates like the ones on Bikini Video or Fashion Walk follow a deliberate structure: a scene-activation opener, the action, the identity-preservation block, then lighting, depth of field and camera movement, and a closing quality instruction. Clearing the box and typing “she walks on a runway” throws all of it away.

The fix: load the closest template, then edit the action clause and the lighting clause only. Leave the preservation block and the closing quality line intact. If you want to see well-formed prompts of this shape in bulk, the Seedance prompt library is a good reference set.

Mistake 7: you iterated at the most expensive tier

Resolution tiers are not small differences. On Seedance 2, a 5-second clip costs 38 credits at 480p, 82 at 720p, and 203 at 1080p. Three failed 1080p attempts is 609 credits — more than double the monthly allowance on the $15 Lite plan.

The fix: block out composition and motion at the cheapest tier that shows you whether the idea works, then spend once at final resolution. The Seedance 2 Fast variant at 480p is 24 credits for 5 seconds, so you get roughly eight test runs for the price of one finished 1080p clip. Full tier pricing is on the pricing page.

Mistake 8: you re-rolled without changing anything

Hitting generate again on an identical prompt is a lottery ticket. Usually you pay twice for the same failure mode.

The fix: change exactly one variable per attempt — the verb, the camera move, the model, or the input image. One variable means you learn something. Three means you learn nothing even when it works.

Mistake 9: you blamed the prompt when the input image was the problem

Image-to-video inherits every flaw in the source. A soft, low-light, filtered photo produces a mushy video no matter how good the prompt is, and occluded hands or cropped feet become motion artefacts.

The fix: start from a sharp, well-lit photo with the whole subject in frame. If you do not have one, clean it first in the image editor, then animate it. To reverse-engineer a look you like, the photo-to-prompt tool will describe a still frame for you.

Mistake 10: you pushed at a filter instead of switching models

Models here are labelled by content-filter strictness. If a swimwear, dance or fitness prompt keeps getting refused, rewording it fifteen times on a strict model burns credits for nothing — switch to a flexible-tier model, which handles that material directly. Explicit and illegal content stays blocked platform-wide regardless of model.

A quick prompt template that works

Subject and action, identity preservation if there is an input image, lighting and lens, camera movement, mood, quality instruction — in that order, roughly forty to eighty words. Longer is not better past about 120 words. Test the pattern on a cheap tier before committing to anything expensive.

Frequently asked questions

Does a longer prompt give a better AI video?

Only up to a point. Forty to eighty words covering action, identity, lighting, camera and mood is the sweet spot. Past roughly 120 words most video models start weighting the tail of the prompt less, so extra adjectives displace rather than add.

Do I get credits back if the generation fails?

Yes. When a job fails on the provider side rather than completing with a result you dislike, the deduction is reversed automatically by the refund path in the credit system. Credits are not refunded for a successful generation you simply did not like — that is why testing at a cheap tier matters.

Can I prompt in a language other than English?

Yes. VdoBloom has an optional prompt translation step that converts a non-English prompt to English before it reaches the model, and passes English prompts through unchanged. Most underlying models are still trained predominantly on English, so translated prompts generally behave better than raw non-English ones.

Why does my subject’s face change halfway through the clip?

Almost always a missing identity-preservation clause, a too-long duration, or a low-resolution source photo. Add the preserve-features sentence, drop to a 5-second clip, and start from a sharper image before you try anything else.

10 AI Video Prompt Mistakes That Waste VdoBloom Credits (and Their Fixes)