· 8 min read
How to create ad creative with AI
The hard part of AI ad creative is not making something that looks good. It is making something that contains your actual product, with legible copy, in the right shape for the placement.
Generic AI imagery is easy and nearly worthless for advertising. What a campaign needs is the specific product, recognisable, with the brand's copy readable in frame, delivered in three or four aspect ratios. Each of those is a solvable problem, and each needs a different part of the toolkit.
Problem 1: keeping the real product in frame
A model asked to invent a bottle of face serum invents a plausible one. It will not be yours. The fix is to supply the product as a reference image and pick a model that genuinely uses it.
- For video, Seedance 2.0 takes up to four subject references and can lock the camera — the combination that makes a clean product turntable possible. Seedance 2.5 holds a single subject across a longer take.
- For images, Seedream 5 Pro and Grok Imagine Quality both edit from a supplied reference. Several other image models accept an upload and ignore it, so check the reference-image row on the model page before relying on one.
Then write the prompt as direction rather than description. The image already says what the product looks like; the prompt should say where it is, how it is lit, and what the camera does. Re-describing the product in words is what causes it to drift away from the reference.
Problem 2: readable copy in the image
Text rendering is the sharpest quality difference between image models right now. Most of them approximate letterforms convincingly at a glance and fall apart on inspection, which is fine for a background and fatal for packaging.
GPT Image 2 is the model to use when words are in the frame — packaging copy, signage, ingredient lists, headline lockups. Ask for the exact string you want in quotes, keep it short, and specify where it sits: the words "COLD BREW" printed in bold sans-serif across the lower third of the can.
Problem 3: one idea, five placements
A campaign needs a 9:16 for Stories and Reels, a 1:1 for feed, a 16:9 for YouTube, and often a wide banner. Cropping one render down to all of them loses the composition every time.
Two better approaches:
- Generate each ratio natively. Run the same prompt at each aspect ratio the model offers. Nano Banana Pro has the widest set of ratios in the image catalog; Grok Imagine covers the extreme wide and tall crops most models skip.
- Generate large, then crop deliberately. Render at high resolution with headroom around the subject and cut each placement by hand. Slower, but it keeps a single composition recognisable across the set.
A production process that works
- Write the brief as a shot, not a mood. "Product on wet slate, hard side light from the left, shallow depth of field, cool palette" beats "premium, modern, aspirational".
- Draft cheap. Nano Banana 2 Lite for images, Seedance 2.0 Mini for video. Find the composition before spending flagship credits.
- Lock the still first. Get the hero image right, then animate it — an image-to-video pass from an approved still is far more predictable than generating video from scratch.
- Render the finals natively per ratio.
- Set type on top. Headline, logo and legal in your design tool, not in the prompt.
What to keep out of the prompt
- Brand names of other companies. Asking for the look of a named competitor's campaign produces derivative work and legal exposure. Describe the lighting and composition you admire instead.
- Real people who have not agreed to it. Use your own talent, with a release, as a reference image.
- Claims. "Clinically proven" rendered onto a pack shot is a regulatory problem, not a design one.
- Everything at once. Prompts that specify twenty things get roughly half of them. Specify the five that matter and let the model handle the rest.
Where the workflow lives
The marketing studio and the editing studio — upload a product and talent shot, generate ad-ready output from a style, then trim, caption, add music and export a finished MP4 — are on the Creator and Studio plans, along with publishing straight to TikTok, Instagram, YouTube and Facebook. Generation and your gallery are on every plan. See pricing for the full comparison.
Models mentioned here
GPT Image 2
OpenAI's image model, and the one to trust with text in the frame
ByteDanceSeedance 2.0
4K AI video with a lockable camera and multi-subject references
ByteDanceSeedance 2.5
Long-form AI video with reference control and native audio
GoogleNano Banana Pro
Google's highest-fidelity Gemini image model, up to 4K
Put this into practice
50 free credits on sign-up, no credit card required.