Our AI video ads look inconsistent from one render to the next. How do we keep the same presenter and product?
AI video ads look different from one render to the next because something is being re-invented each time: a presenter described in words, a product drawn from a prompt, or settings that change with whoever pressed the button. Fix it by turning each into a fixed input. Use one approved avatar with its face and voice locked, render every product shot from a real image of the product, save the settings as a shared preset, and make variants from shots already rendered.
Below are the causes in the order they usually bite, the fix for each, a check to run before an ad ships and what it costs to skip it. Nagent's Alpha, our AI video teammate, is built around these rules, and it has its own section near the end.
Why do AI video ads change from one render to the next?
Generative video models do not remember your last render. Each generation starts from the inputs it is given, and anything left to the model is decided again. Six causes account for most drift:
The presenter lives in the prompt. "A woman in her thirties, warm smile, kitchen" gives a new face every time. Even small wording changes move the result.
The product is described, not shown. A model asked for "a white bottle with a green cap" invents a bottle. Labels, cap shapes and colours shift between renders.
Different models or modes per shot. Text to video, image to video and lip sync models each render faces slightly differently. Stitching clips from several models makes the seams visible.
Settings change between teammates. One person renders at a fast draft quality in 4:5, another at a premium quality in 9:16, with a different style note. The brand looks like two brands.
The voice is not fixed. If the voice is picked per job, or a model generates speech from a still in its own voice, the presenter sounds different even when the face holds.
Re-rendering whole videos for small edits. Changing one line by regenerating the full ad re-rolls every shot, including the ones that were right.
How do you keep the same presenter in every render?
Make the presenter an asset you pick by name, not a description you type.
Use a custom or saved avatar. Every serious tool supports one. HeyGen says you can reuse the same avatar, voice and brand look across hundreds of videos (HeyGen). Creatify builds custom digital twins from 2 to 5 minutes of video footage (Creatify). Synthesia makes personal avatars from a photo or short video (Synthesia).
Start from one master still. Choose one well-lit, front-facing image as the face every version keeps. Do not swap it for a "better" photo halfway through a campaign.
Fix the voice. Clone or design one voice for the presenter and pin it to the avatar, rather than choosing per video.
Write the look down. Record wardrobe, setting and delivery in words alongside the avatar, so the next person does not improvise them.
Get consent in writing for a real person. If the presenter is a founder, employee or creator, you need their release before you build a twin. Synthesia requires a live on-camera consent recording that cannot be skipped. Creatify's own guide says verbal permission is not enough and lists what to document, including how long the likeness can be used and what happens if the person wants the avatar retired.
Keep one presenter per concept. Consistency inside a campaign is good; the same face on every ad you run tires fast. Rotate presenters between concepts, not between variants of one concept.
How do you keep the product accurate in every shot?
The product is where inconsistency costs most, because an ad that does not match the shelf loses trust and can mislead.
Always render from a real product image. Upload a clean packshot and make it the reference for every product shot. HeyGen's product placement tool, for example, asks for a product image of 720p or higher with the item front and centre (HeyGen).
Keep one approved image per product and angle. Store it in one library, name it by SKU, and stop people pulling old photos from a shared drive.
Avoid asking one clip to do two jobs. A single generated clip usually starts from one still. If a shot needs both the presenter and a close-up of the pack, plan them as separate shots and cut between them.
Keep on-screen words out of the image model. Supers, prices and the offer should be set as text in the edit, not drawn by the video model, which can misspell them.
Check the label at full size. Look at the frame where the product is largest. Colour, cap and logo drift show up there first.
How do you stop settings drifting between teammates?
Save the decisions once and make them the default.
Create a preset per placement. One for Meta Reels at 9:16, one for feed at 4:5, each with its length, quality and style note fixed.
Lock the brand kit. Logo, colours and end card come from one brand file, not from whoever has the latest version.
Fix the model choice. Pick one quality tier and one model style for the campaign. Use fast drafts only for internal review, never for the final cut.
Name the owner. One person approves changes to a preset. Everyone else uses it.
Should you re-render or make a variant?
Make a variant whenever the change is not to the shots themselves. A new hook line, a different call to action, a fresh end card, captions, another language or a new ratio should reuse the shots already rendered. Re-rendering the whole video for these re-rolls the presenter and product and charges you again for shots that were right. Ask any vendor how its variants work before you buy: does a call to action swap render anything at all?
How do you check consistency before an ad ships?
Run this five-minute check on every cut:
Pause on the presenter in the first and last shot. Same face, hair, wardrobe and setting?
Listen to the first and last line. Same voice, pace and accent?
Pause on the largest product frame. Same label, colour, cap and logo as your packshot?
Read every word on screen against the brief. Any line, figure or price nobody wrote?
Watch with sound off. Does the lip sync hold where the presenter speaks?
If two of five fail, fix the inputs rather than re-rolling the render and hoping.
What does inconsistency cost?
Re-renders. Each re-roll spends credits or minutes on shots you already paid for.
Ad review and trust. A product that does not match the real one can read as misleading, and a viewer who notices a changed face stops trusting the ad.
Lost learning. If every variant changes the presenter, product and hook at once, you cannot tell which change moved the result.
Likeness risk. Rendering a real person without a current release exposes the brand, and an expired release means every video with that face needs retiring.
Where does Nagent fit?
We make Nagent, so here is what Alpha does about each cause, and where it stops.
Presenter as a locked persona. In Alpha, a presenter is a persona with a master still, a voice, wardrobe, setting, delivery and a disclosure line. When a persona is approved, by another admin of your workspace or by Nagent, never by the person who asked, its still, voice and disclosure are locked. Every job that casts it starts from them, and a persona that cannot be cast today is not offered.
Real people only with a release. A digital twin renders only with a written release on file that says what it covers, where and until when. Revoking it blocks new renders at once and flags every video made with it so it is not reused.
Product from the exact pack. Any product shot starts from an image of the real product chosen from your RUPA image library, so the pack, label and colour match. A storyboard with a product shot waits for that image before it can go for approval.
Settings remembered for everyone. "Remember for" keeps a preset's settings for the whole workspace after a confirm card, and Brand Lock sends your brand pack with every render.
Variants that reuse shots. A hook swap, call to action swap, end card refresh, resize, captions or a new language keeps the shots already rendered and renders again only what changes.
A check on every cut. PROOF watches and listens to each cut: it checks that every face in the presenter's shots is the same person, that the mouth moves with the words, and that the words spoken match the storyboard.
What it does not do: PROOF asks for a fix rather than holding a cut back on these checks, so a person still reads its report. A persona made from a still may speak straight from that still in the model's own voice rather than its calibrated one, which is a known source of voice drift. Alpha's studio is labelled Beta, presenter shots render only where those lanes are connected for your workspace, and a render takes 30 minutes to 2 hours. A rendered video is about 700 credits as a rough guide; see pricing.
For why the quote should come before the render, see why AI video tools should quote before they render.
Frequently Asked Questions
Why does my AI avatar look slightly different in every video?
Usually because it is rebuilt from a description or a different photo each time, or because different shots use different models. Pick one saved avatar with one master still and one voice, and keep the model and quality tier fixed for the campaign.
How do I stop AI from changing my product's label or colour?
Never ask the model to draw the product from words. Give it a clean, high-resolution image of the real product as the reference for every product shot, keep one approved image per angle, and set prices and claims as text in the edit.
Can I use a real person as a consistent AI presenter?
Yes, with their written consent covering where and how long the likeness is used, and a clear way to retire it. On Nagent, a twin also needs approval from your workspace and from Nagent before its likeness is bound.
Is there an AI video tool that shows a storyboard and a price before rendering?
Alpha does: each request comes back as a storyboard with a credit quote, and nothing renders until a person approves both. The quote is fixed at approval.
Do variants change the presenter?
They should not. A variant that only changes the hook, call to action, end card, captions, language or ratio should reuse the rendered shots. Check how your tool handles it; on Alpha, variants keep the shots already rendered.
See how Alpha locks a presenter and a product for your brand: book a Nagent demo.