AI Video for Amazon Ads: The 16:9 Problem

Quick answer:Amazon's ad-specs page for Sponsored Brands video lists the aspect ratio as "16:9 (square pixel) only", 6 to 45 seconds, 500MB or less, MP4 or MOV. Its moderation guide separately bans black bars, so you cannot letterbox a vertical render into a horizontal frame to satisfy the first rule. That collides with how AI video tools are built: they default to 9:16, and on several image-to-video models the aspect-ratio dropdown is ignoredthe moment you supply a starting image, because the image's own ratio wins. So an Amazon video is not a setting you flip at export. It is a 16:9 reference frame chosen before you generate, a crop-to-fill rather than a pad, and, past 15 seconds on a scene model, either a stitch or a switch to an audio-driven model that has no duration control at all.

The spec, as Amazon writes it

This is the Sponsored Brands video page, read in August 2026. It is worth quoting rather than summarising, because the third-party guides that rank for these queries paraphrase it in ways that lose the conditions.

Two of those rows quietly rule out most AI output as it comes off the model. The frame rate list stops at 30fps. And the video profile row asks for Main or Baseline, while ffmpeg's libx264 encoder defaults to High. High-profile files are widely accepted in practice, but if a creative is rejected and the reason is unhelpful, re-encoding with an explicit Main profile is a cheap thing to rule out. Run ffprobe on the file before you argue with moderation.

Vertical exists, but narrower than the blogs suggest

Search for Amazon video specs and you will find plenty of pages saying Amazon takes 9:16 now. Half true. Amazon announced vertical video creatives for Sponsored Brands, and the announcement is specific about where they run: the drive page visit goal with one to three advertised products, and the grow brand impression share goal with zero to three, in both cases with a Store as the landing page. Its availability section does not print a total; it lists 21 marketplaces, three in North America, one in South America, ten in Europe, three in the Middle East and four in Asia Pacific.

The ad-specs page itself does not mention vertical at all. It carries one specification table and one aspect ratio line, and that line says 16:9 only. So if you are running the ordinary Sponsored Brands video that lands on a product detail page, horizontal is still the assumption you should build to. Vertical is a second creative you make deliberately for a Store-landing campaign, not a substitute for the first.

Why you cannot just pad the frame

The obvious workaround for a library of 9:16 UGC is to drop each clip into a 1920x1080 canvas and fill the sides with black. Amazon closes that door explicitly. Its Sponsored Brands and display ads moderation guide says videos "must fill the available space" and "must not have black bars on any side, which means letterbox and pillar-box formats are not permitted". The same guide lists "black or blank frames at the start or end of videos" among its common rejection reasons, alongside blurry or pixelated visuals.

This is a real fork in an export pipeline, not a preference. Ours has both branches. The render route takes a target aspect and a fit mode, and the two modes are different ffmpeg filter chains:

• Pad scales to fit and fills the short axis with black. Every pixel of the original survives. Amazon rejects the result.
• Crop to fill scales to cover and centre-crops to the exact canvas. Nothing is padded, and pixels at the top and bottom of a vertical source are thrown away. This is the Amazon-safe branch, and it is our default.

Crop-to-fill has its own cost, and Amazon names it: when adjusting the video creative, do not crop out the product or any relevant information. Take a 1080x1920 talking-head clip, centre-crop it to 1920x1080, and you keep a horizontal band across the middle of the frame. If the product was held at chest height, it survives. If the presenter held it up near their face, or your caption sat in the lower third, it does not. That is the actual reason to generate horizontally instead of converting: not fidelity, framing.

The dropdown does not do what you think

Here is the part almost nobody writes down, and it comes straight out of the model schemas we dispatch against. On several image-to-video models the aspect ratio you pass is ignored when you supply a starting image. The image's ratio wins.

Four models carry that note in our dispatch layer, in the schema's own words: Happy Horse 1.0 and Happy Horse 1.1 (aspect ratio "ignored when a single image is provided"), Seedance 1.5 Pro ("ignored when an image is used") and Kling 2.6 ("ignored when start_image is provided"). Grok Video is the same trap by a different route: its schema has an explicit autovalue that follows the input image's native ratio. Kling 3.0 is a softer case, noted as "only required without a start frame". For Seedance 2.0 and Veo 3.1 we have no such note either way, which is a reason to test rather than a reason to assume. Four families, in practice:

• Aspect parameter, documented as overridden by your image. Happy Horse 1.0, Happy Horse 1.1, Seedance 1.5 Pro, Kling 2.6, and Grok Video via auto. Set 16:9 and supply a 16:9 first frame, or the setting does nothing.
• Aspect parameter, behaviour with an image not documented. Veo 3.1 Fast, Kling 3.0, Seedance 2.0 (which we run on BytePlus rather than Replicate). We send the ratio you pick. Whether the first frame overrides it is not written down in the schemas we work from, so render one test clip before you commit a batch.
• No aspect parameter at all.OmniHuman 1.5, VEED Fabric 1.0 and Pruna Avatar are audio-driven: you give them a still and either a voice track or a script, and the output inherits the still's shape and the speech's length. The reference image is the aspect control.
• A two-value enum. Sora 2 takes portrait or landscape rather than a ratio string. Do not build an Amazon workflow on it regardless: OpenAI retires Sora 2 and Sora 2 Pro on 24 September 2026.

Practical consequence: the first asset in an Amazon pipeline is a horizontal reference frame. Generate or shoot the product still at 1920x1080 before you touch the video model, and the aspect problem stops being an export problem.

Duration is the second wall

Amazon's floor is 6 seconds and its own recommendation is 20 or less on the spec page, while its Sponsored Brands video guide says 15 to 30. Against those numbers, here is what a single clip can actually be. Durations and credit costs below are ours, taken from the live pricing table rather than a vendor marketing page.

Read that against Amazon's window and the shape of the problem is obvious. Every model with a duration control stops at 15 seconds, which is the bottom of Amazon's own preferred range. Anything longer from a scene model is two or three clips cut together.

The three audio-driven models go further, and it is worth being precise about why, because a widely repeated version of this claim is wrong. There is no 30-second capability ceiling. 30 seconds is simply the longest row in our published price table for OmniHuman 1.5 and VEED Fabric 1.0, and that table is not what the app charges for these three. They have no duration input at all, so the charge is calculated per spoken second from the script, and our shared estimator runs to 120 seconds. Where each model actually stops is a separate number, and they differ: our registry caps OmniHuman 1.5 at 30 seconds because its schema requires an audio track under 35 seconds, VEED Fabric 1.0 at 120, and Pruna Avatar bills to 60 because the model truncates speech around there. The voice track decides how long the video is, and therefore what it costs. For a 6-to-45-second Amazon creative none of those ceilings binds.

The 720p rows are not a compromise, incidentally. 1280x720 is one of Amazon's three accepted dimensions, so a 720p export is a fully compliant Sponsored Brands video. 4K is accepted too, at 3840x2160, though the 500MB ceiling gets closer at that resolution, and the moderation guide asks for larger on-screen text at 2160p: 100pt minimum, 200pt preferred, against 50pt minimum and 100pt preferred at 720p and 1080p.

If you are choosing between models on cost rather than length, we keep the full per-second comparison in the AI video model cost benchmarks, and the model picker filters the catalogue by what you actually need from a clip.

Two workable Amazon builds

The 6 to 15 second single-clip build.One horizontal reference frame, one scene model, no stitching. Veo 3.1 Fast is the obvious first reach and it is not the cheapest: at 6 seconds it is 180 credits, while Grok Video at 10 seconds and 720p is 140 and Seedance 1.5 Pro at 8 seconds and 1080p is 160. All three clear the 6-second floor, and 1280x720 is one of Amazon's three accepted dimensions, so the Grok route is fully compliant rather than a downgrade. Sora 2 at 8 seconds is also 160, but it retires on 24 September 2026, so do not start there. At the top of the range, Happy Horse 1.1 or Kling 3.0 at 15 seconds lands inside Amazon's recommended range in a single generation. This is the build to default to, because every cut you avoid is a continuity risk you avoid.

The 20 to 30 second spokesperson build. A 16:9 still of your presenter, plus a voice track for OmniHuman 1.5 or VEED Fabric 1.0, or a typed script for Pruna Avatar, which runs its own text to speech. Length comes from the speech, so a 26-second read yields a 26-second video with no seam, and the credit cost scales with it: 26 spoken seconds is 849 credits on OmniHuman, 797 on Fabric and 478 on Pruna Avatar. The trade is that you are locked to one shot: no camera move, no cutaway, just the presenter. For a product with a story to tell that is often fine, and it is the same production logic behind audio-driven lip sync generally. If you want cutaways, generate the b-roll separately on a scene model and cut it in.

One honest note on our own product. Six models are blocked during the free trial: Kling 3.0, Seedance 2.0, Sora 2, Sora 2 Pro, OmniHuman 1.5 and VEED Fabric 1.0. The trial also caps any single render at 250 credits. Between them, that makes the long spokesperson build a paid-plan workflow.

Pruna Avatar is the one to be careful with, because its static price row is misleading and we would rather say so than let you find out from your balance. That row reads "~20s for 180 credits", but it is a fallback anchor, not the price. Pruna has no duration input, so the app bills it per spoken second at 18.4 credits a second. A 20-second render is 368 credits, which is over the 250-credit trial ceiling, and the render is refused rather than trimmed. Under that ceiling Pruna gives you 13 spoken seconds, which at the 2.3 words-per-second rate our estimator uses is 29 words of script.

So for testing the Amazon path during the 3-day trial, the longest combination that fits under 250 credits on each unblocked model is Veo 3.1 Fast at 8 seconds and 1080p (245), Seedance 1.5 Pro at 12 seconds and 1080p (245), Grok Video at 10 seconds and 720p (140), Happy Horse 1.1 at 5 seconds and 1080p (180), Kling 2.6 at 5 seconds (125), and a Pruna Avatar script of 13 spoken seconds or less. Every shorter tier on those models fits too. The first three, plus Pruna, clear Amazon's 6-second floor. The two 5-second tiers do not, so treat those as look tests rather than as uploadable creative.

A pre-upload checklist

• Reference frame is 16:9 before generation, not cropped afterwards.
• Export fills the canvas. No pad, no bars, no blurred-edge fill.
• Product still readable after the crop. Check the top and bottom thirds specifically.
• Length between 6 and 45 seconds. Aim 15 to 20 unless you have a reason not to.
• Dimensions exactly 1280x720, 1920x1080 or 3840x2160.
• Under 500MB, MP4 or MOV, H.264 or H.265, progressive, 30fps or lower.
• No black or blank frames at the head or tail. Trim them.
• On-screen text at 50pt minimum for 720p and 1080p, 100pt minimum at 2160p.
• Captions burned in. Sponsored Brands video autoplays muted, so the first two seconds have to work silently.
• Audio language matching the marketplace. Amazon checks this.
• Separate from all of the above: the AI-labelling metadata, if your creative contains a photorealistic AI-generated person. New York's synthetic performer law and Amazon's contains-synthetic-performer tag cover which keyword goes in which field, and why it has to be written after your last re-encode rather than before it.

What we could not verify

Everything above about ads comes from public Amazon Ads pages. The specs for listing videos, the ones on a product detail page or inside A+ content, live in Seller Central behind a login. They are widely reported as matching the ad specs, but we could not check that against an Amazon-hosted page and will not assert it from a source we cannot open. If you are uploading to a listing rather than buying media, read the requirement in Seller Central itself.

Sources

• Amazon Ads, Sponsored Brands video, ad specs and guidelines (aspect ratio, dimensions, length, file size, codec, profile, frame rate, audio, bars)
• Amazon Ads, A complete guide to Sponsored Brands and display ads moderation (letterbox and pillarbox ban, rejection reasons, on-screen text minimums)
• Amazon Ads, Sponsored Brands video introduces vertical video creatives (which campaign goals accept 9:16, and the 21 marketplaces its availability list names)
• Amazon Ads, Get started with Sponsored Brands video (both formats supported, 15 to 30 second guidance, muted autoplay, captions)
• Amazon Ads, Amazon Video Ads specs (the separate out-stream format, for contrast: 1920x1080 recommended, up to 3 minutes)
• Forbes, "Amazon Requires Sellers To Label AI-Generated People In Listing Images", 25 July 2026

All six were read directly in August 2026. Model ceilings, aspect-ratio handling and credit costs come from our own dispatch and pricing tables, which change when a vendor changes a schema or a price.

Selling on Amazon specifically? AI UGC ads for Amazon FBA sellers covers the wider workflow, and how to make product demo videos with AI covers the creative itself. UGC Vids AI shows the credit cost of every model, duration and resolution before you generate, and exports at 16:9 with a crop-to-fill that Amazon will accept. Plans start at $49/month for 5,000 credits. Free for 3 days.