Why do AI ads all look the same, and how do you make yours look like you?

Every image model has a most likely answer for a scene it wasn't told much about. Ask for "a premium skincare bottle on a clean background" and the model has to choose the background color, the direction of the light, the lens, the depth of field and the finish by itself. It chooses what is typical, so every brand that writes that prompt gets a close cousin of the same picture. Google's own advice for prompting its image models is to describe the scene instead of listing keywords (Google for Developers, August 2025), and "premium, clean, modern" is the kind of prompt that advice warns against.

Three habits make it worse. Mood boards get pulled from the same trend references, and when the board's objects are copied into the scene, every brand that used that board gets the same set. Headlines get drawn by the image model, so the letterforms are the model's. And the first acceptable render ships, defaults and all, because nobody made a second or a tenth.

The volume makes the sameness easy to spot. Meta's engineering team reported that more than a million advertisers made over 15 million ads in one month with Meta's own generative AI tools, and built its Andromeda ad retrieval system expecting the number of ad creatives to keep growing (Meta Engineering, December 2024, checked October 2026). All of those ads started from the same set of tools.

One reason is a free tool TikTok built for exactly this job. TikTok's Symphony Creative Studio generates an ad video from a few product details or a product URL, offers stock digital avatars as presenters, and translates and dubs the voiceover, and TikTok opened it to every logged-in TikTok for Business user in June 2025 (TikTok for Business, June 2025, checked October 2026). When many advertisers feed a product page into the same generator and pick from the same avatars, the layouts, pacing and faces repeat. The prompt problem from the top of this page applies to video too: a product URL tells the model what to show and nothing about how your brand lights, frames or cuts it.

Viewers are also told which ads are AI. Videos made in Symphony Creative Studio are labeled "AI-generated" automatically (TikTok for Business, November 2024, checked October 2026), and ads with realistic AI-generated or AI-altered images, video or audio need the AI-generated disclaimer, which TikTok places under the in-feed ad (TikTok Ads Manager Help, updated September 2025, checked October 2026). The label stays either way, so the only lever left is whether the ad under it looks like your brand or like everyone else's.

Does looking the same make them less effective? TikTok hasn't published performance data on that, and the NIQ study above measured attitudes and memory, not sales. What it does tell you is that people notice AI ads and remember them less, and a stock avatar that a viewer saw in another brand's ad yesterday is the fastest way to be noticed for the wrong reason.

How to make an AI video ad stand out

Start from a still that already passes the five checks above, then animate it, instead of asking a video model to invent the scene from a URL. The first frame carries your hex codes, light and lens, and the model only has to move it. Expect more waste once things move: on our hand-built brownie job, food shots with motion took 12.0 renders per keeper, against 3.2 for packaged gummy ads. Overs makes stills only, but it writes a text motion prompt you can hand to a video model.

Skip the stock presenter. Use a person who is yours (a creator you cast, your founder, your customer) or let the product carry the ad, and if you need talking-head energy, UGC-style ad creatives covers the formats. Put captions and on-screen text in during the edit, in your own fonts, for the same reason you keep type out of the image model. And change the idea between variants: a new hook, a new angle on the product, a new setting. Swapping the background tint across five cuts gives TikTok five copies of one ad. If you're deciding where video earns its cost at all, static vs video ads compares the two.

NIQ showed low- to high-quality AI-generated ads to more than 2,000 people and measured brain activity in about 150 of them. Viewers identified most of the AI-generated ads, rated them more annoying, boring and confusing than traditional ads, and even the high-quality ones produced weaker memory activation (NIQ, December 2024, checked October 2026). The study measured attitudes and memory. It did not measure click-through or cost per purchase, so treat it as a warning about recall: an ad people don't remember does less for the next ad they see from you.

Sameness also wastes testing budget. Five renders that differ only in background tint test one idea, and the results tell you little about which angle to make next. If you're planning volume, how many ad creatives you need and our ad creative testing framework cover how to spread a batch across real differences.

Run these on the last batch you shipped. They take about ten minutes and need nothing more than an eyedropper tool.

1. Cover the logo. Could the ad run for your closest competitor without anyone noticing? If it could, the brand lives only in the logo.
2. Name the light. Say where the key light comes from and how hard it is. If you can't, the model chose it.
3. Sample the colors. Pull the background and shadow colors with an eyedropper and compare them with your palette's hex codes.
4. Look at the type. If the headline letterforms came out of the image model, they aren't your fonts.
5. Line up 20 renders. Put the last 20 side by side. Repeated backgrounds, camera angles and depth of field mean the prompt is carrying defaults.

If an ad fails only the first check, the look is right and the scene is generic. Fix the scene before you touch the style.

Each row below is a choice every image needs. Leave it out of the prompt and the model makes it. The right-hand column shows how our kept prompts made the same choice as a brand decision.

The prompts behind the photos we kept from four campaigns built by hand in 2026 ran about 140 to 200 words on Nano Banana models and about 300 on GPT Image. The Nano Banana ones followed one order:

1. The subject and a pointer to the reference image.
2. The background, named with a hex color.
3. Light, described as something falling on a surface.
4. The palette, as names plus hex codes.
5. One sentence of camera and lens.
6. A separate sentence about grain.

Here is one, from an apparel close-up on a model:

A tight chest crop of the model shown in the reference image wearing the exact striped henley pullover from the reference, two wood-grain henley buttons visible at the placket, drawstrings resting asymmetrically across the chest, subtle woven patch on the left breast, subtle natural breathing shaping the drape. Background is a softly out-of-focus Sea Foam White #F2F5F8 seamless cyclorama falling completely away to push all attention onto foreground fiber. Hard raking natural daylight enters at a 45-degree angle across the chest, carving the herringbone weave and revealing the muted dusty mauve and Faded Sky Blue #B0C4DE threads inside the stripes. Shot on a 100mm macro lens at f/4, hyper-crisp focus riding the button and the adjacent thread structure, 1:1 square crop centered on the placket. Digital medium-format color profile graded toward Kodak Portra 400, clean near-zero grain, balanced contrast preserving honest shadow detail inside the fabric folds, natural saturation emphasizing the washed, lived-in heavyweight cotton texture.

A kept Nano Banana 2 prompt, about 150 words. Lightly edited: brand names removed.

The color has a code, the light has an angle, the lens has a focal length and an aperture, and the grain is named. Those four details are what make it look like one brand.

On our brownie job, all ten kept prompts asked for heavy film grain, and nine carried the same sentence: "Heavy visible film grain throughout." A person added it by hand, prompt by prompt. Only 14 of the job's 120 prompts carry it, so nine of the ten keepers share a sentence that 106 of the job's prompts left out.

Pull your best existing images and write down what they share: the palette as hex codes, where the light comes from and how hard it is, the lens, the grain, and how things are styled. That list is your visual DNA, and it goes into every prompt. Overs' guide to visual DNA has a template for writing it down.

Then give those rules the final say. In Overs, brand rules are applied last in the ten-step run, and the quality check only checks the work against them. It never rewrites the art direction.

If several brands copy the objects from the same trend images, they end up with the same scene. Take the light, the mood and the pace from a board and leave the objects behind. Overs reads a board of up to twelve images plus a note for its feeling and applies that to your product in your brand's style, and a dial decides whether your brand's own look or the board leads.

An image model drawing a headline picks the letterforms itself. Set type in a separate pass, in your own fonts and grid. Overs does this in its own typography step, setting a headline, subline and button on a finished photo and keeping the clean version. On the ad side, what makes a good static ad covers layout and hierarchy, and static ads for Meta covers the placements and sizes the type has to survive.

In our four hand-built campaigns, 361 renders produced 45 keepers: 3.2 renders per keeper on gummy supplement ads, 3.6 on trousers, 12.0 on brownie ads and 13.6 on a surf campaign carried across three scenes. Somebody compared each render with the brand and threw away 316 of them.

The ratio tracks how much the image has to invent. A packaged product on a set background gives the model little to guess. A food shot with motion, or one model carried through three outdoor scenes, gives it a lot, and each guess is another chance to drift from the brand.

In Overs, those decisions stay on record. Approvals and rejections with reasons become brand memory, and a note that fixed an approved photo is remembered. Brand Watch reads the brand's recent Instagram posts so the look stays current: on request at most once a month on the free plan, weekly on Pro and daily on Team (Overs pricing, checked October 2026).

First, the written brand rules before the first render: hex codes, light, lens, grain and styling, in a document you can keep. Second, one kept prompt from your job, so you can see whether your brand's decisions are in it or whether it reads like "premium, clean, modern". Third, the rough number of renders thrown away per keeper. An agency that can't tell you is probably shipping first acceptable renders.

100 Creatives starts every job with a brand intake on visual language, model preferences, colorways and export specs, and returns creative batches within 48 hours. How to brief an agency for AI creative has a template for the intake, and how to review AI creative from an agency covers what to check when the batch arrives.

Overs was built on the process behind those 45 keepers, and it is what our team makes its images in today. It records your visual DNA with hex codes, light, lens, grain and styling, applies your brand rules last, and shows the shot plan and the cost before anything renders. For anything worn or held, you approve a character sheet first, so the same model appears in every photo. Overs lists renders at about $0.03 to $0.21 per photo depending on the model (Overs pricing, checked October 2026), so at our average of 8 tries per keeper, a usable photo costs roughly $0.24 to $1.68 in AI fees.

Overs is our sister product, made by 100 Creatives' founder. Try Overs free at www.overs.studio. The free plan makes 40 photos a month and Pro is $29 a month, with the AI billed on your own OpenRouter account and no markup from Overs.