What are the ways to make a product ad with AI?
A product ad can be made with AI in four ways: a general image generator, a platform's own creative tools, a dedicated product ad tool or a designer who uses AI for the base image. All four depend on the same input, a real photo of the product, because the ad has to show the item that ships.
Google Ads has built-in options too. Its image enhancements use Google AI to create additional versions of an image, for example by adding text overlays and adjusting aspect ratios, and Google notes that images are not always enhanced.
| Route | Example tools | Text handling | Best for | Main risk |
|---|---|---|---|---|
| General AI generator | ChatGPT, Gemini, Midjourney | Drawn into the image | Concepts and mood images | The product and the letters can change |
| Ad platform tools | Google Ads image enhancements | Added by the platform | Extra versions of an approved image | Less control over each result |
| Dedicated product ad tool | Herofab Ad creative | Set as real type after the image is made | Several formats from one product photo | The scene still needs a product check |
| Designer or agency | A design tool with AI features | Set by hand | Campaigns with a strong brand system | Cost and turnaround |
Make a product ad with Ad creative
Which image formats do Meta, Google, TikTok and Instagram ads use?
Three ratios cover most image placements: 1:1 and 4:5 in feeds, and 9:16 in stories. Google also needs a horizontal 1.91:1 image. The table gives the sizes each platform publishes; check the preview in the ad manager before launch, because placements and limits change.
Google's figures were read on 31 October 2026. TikTok's come from the Herofab guide checked on 12 October 2026. Meta's and Instagram's ratios come from Sprout Social's guide updated 15 October 2026 and Kive's 14 October 2026 summary of Meta's Ads Guide, not from Meta's own pages: confirm pixel sizes and safe zones in Ads Manager. Keep key content in the centre: Google's Performance Max guidance asks for it in the centre 80% of the image.
| Channel | Placement | Ratio | Size or limit |
|---|---|---|---|
| Meta and Instagram | Feed | 1:1 or 4:5 | Confirm in Ads Manager |
| Meta and Instagram | Stories and Reels | 9:16 | Confirm in Ads Manager |
| Google Demand Gen | Square | 1:1 | 1,200 × 1,200 recommended; 300 × 300 minimum; 5 MB |
| Google Demand Gen | Horizontal | 1.91:1 | 1,200 × 628 recommended; 600 × 314 minimum |
| Google Demand Gen | Vertical | 4:5 and 9:16 | 960 × 1,200 and 1,080 × 1,920 recommended |
| Google Performance Max | Required images | 1.91:1 and 1:1 | 4 or more of each, up to 20; 4:5 is optional |
| TikTok | Carousel ads | Vertical, square or horizontal | 720 × 1,280, 640 × 640 or 1,200 × 628; 2 to 35 images; music required |
What does TikTok ask of a carousel image?
A TikTok carousel takes 2 to 35 uploaded JPG or PNG images, with music, one caption and one link. TikTok suggests 100 KB or less per image, so a large file has to be compressed first. Its policies do not allow blurry visuals or incomplete and illegible text.
How do I make an Instagram post and story from one product photo?
One product photo becomes an Instagram post and a story by reframing it, not by cropping it. Start from a portrait or square image, then extend the scene above and below to reach 9:16. A crop from a square to a tall format cuts the product; an extension adds wall, sky or floor around it.
Stories need clear ends. The interface usually places an account name and a progress bar at the top and a reply field or a link at the bottom. Keep the face, the product and any text in the middle band. For a feed post, a 4:5 image takes more screen height than a square.
An organic post has more freedom than a paid ad. An Instagram Shop image does not: Herofab's Facebook and Instagram Shops guide lists text overlays, calls to action, promo codes and watermarks as not allowed on catalogue images, which should use one ratio, 1:1 or 4:5.
Reframe an image for a story with Extend frame
Should the text be inside the image or in the ad fields?
Put the text in the platform's ad fields whenever the placement offers them, and add it to the image only where the rules allow. Fields stay readable at every size, can be tested without new artwork, and do not break when the platform resizes the image. Google's rules vary by format, so the same artwork may pass in one campaign and fail in another.
Lettering drawn by an AI generator is the second problem. Image models misspell and distort letters, and Shopify's guide tells ChatGPT users to review generated text for spelling. Generate the picture without text and set the words as type afterwards, in any design tool, so they read exactly as written.
| Google placement | Text on the image | Google's wording, read 31 October 2026 |
|---|---|---|
| Search image assets | Not allowed | Added text, logos or graphic overlays are disapproved |
| Responsive display ads | Advised against | “Avoid inserting text on top of an image”; a best practice, not a requirement |
| Performance Max | Allowed with limits | At least one image without overlays is recommended for each aspect ratio |
| Merchant Center product images | Not allowed | Post-production text and marketing callouts cause disapproval |
What should I test in AI product ads?
Test one change at a time, starting with the change most likely to alter the result: the type of image. AI makes variations cheap, which is a risk as well as a benefit. Ten variations that differ in five ways show which ad won but not why.
Judge a test on cost per purchase, and check the return rate afterwards. An image that attracts clicks but overstates the product can win on click-through and lose money. Run each test for its planned length instead of stopping at the first lead.
Google's platform tools support this. Custom experiments split a campaign's budget or traffic, and Google advises at least 10,000 users in an audience list when using a cookie-based split, as smaller lists may give less accurate results. On asset counts, Google recommends uploading 3 of each aspect orientation for Demand Gen, 4 or more unique images for Performance Max, and 5 to 10 images per aspect ratio for responsive display ads.
| Test | Version A | Version B | What it tells you |
|---|---|---|---|
| Image type | Product alone | Product on a model | Whether context sells this item |
| Setting | Plain studio | Lifestyle scene | Whether a scene earns more attention |
| Headline | A benefit | A price or an offer | Which message moves buyers |
| Format | Square 1:1 | Portrait 4:5 | Which uses the feed better |
| Style | Minimal | Seasonal | Whether a seasonal look helps or dates the ad |
Do AI product ads need to be disclosed?
AI product ads often need disclosure, and the rule depends on the platform and the place. Google's advertising policy says rules in the EU, India and New York require disclosures or labels on ads with certain AI-generated or edited assets, and that using its AI label setting does not guarantee compliance. TikTok requires disclosure of content that is completely AI-generated or significantly edited by AI. New York also regulates photorealistic AI-generated people in advertising.
The current details, with sources, sit in the Herofab guides to AI image rules, and this page does not repeat them. Keep the AI metadata in each file, and treat any label set in the ad manager as the advertiser's own task. This is not legal advice.
AI product image rules for 2026
Can I make a fashion lookbook with AI?
A fashion lookbook can be made with AI when each style exists as a sample that can be photographed. A lookbook is read as a sequence, so one model should run through the whole collection. Location images set the mood of the season, and studio white images give a buyer an overview of the line in one pose and one ratio.
A lookbook differs from an ad in its reader. A retail buyer studies the line, so styles have to be shown accurately. A note that the images are AI-generated, placed where the buyer will see it, can avoid a dispute later.
AI lookbook photos for wholesale
Can AI make a video ad from a product photo?
Yes, video generators can animate a product photo. Herofab does not create video, so a video ad needs another tool.
When are AI product ads the wrong choice?
AI product ads are the wrong choice when the image would claim something the product has not shown. Performance claims, such as a before-and-after result or a waterproof test, need real evidence. A generated scene cannot prove them. Size, fit and scale must not be exaggerated by the setting.
A real shoot also wins in three common cases: a brand campaign where the people are part of the message, a reflective or transparent product that generators handle poorly, and an unreleased product with no finished sample. Google's display guidance goes further on realism, recommending physical settings with natural shadows and lighting over digital composite backgrounds. A clear, real product photo with few edits is often a strong first test.
How to check an AI product photo before you publish it
Checklist
- The ad shows the real product: colour, shape, logos and details compared.
- Every word checked letter by letter, in every format.
- Key content sits in the centre, clear of interface areas.
- At least one text-free image per ratio is ready for Google.
- One change per test, with a planned length and a purchase metric.
- The AI metadata is still in the file, and any platform label is set.
- No claim about results, size or fit depends on a generated image.
Sources
- Google Ads Help — Image assets for Demand Gen campaigns
- Google Ads Help — Image assets for Performance Max
- Google Ads Help — Best practices for responsive display ads
- Google Ads Help — Image assets for Search campaigns
- Google Ads Help — Creative image enhancements
- Google Ads Help — Set up a custom experiment
- Google Advertising Policies Help — Image assets format requirements
- Google Merchant Center Help — Text on image
- Shopify — ChatGPT Images guide
- Sprout Social — Social media image sizes guide
- Kive — Meta ad sizes 2026
Checked 31 October 2026. Rules change; the official pages are the final reference.