AI Image-to-Video for Product Ads: Tools, Prompts and Cost Per Finished Ad

An AI image-to-video product ad costs about $14 to $42 per finished 15-second vertical ad in October 2026, built from published per-second rates and a realistic four-to-one take ratio. The list prices look far cheaper: $0.05 to $0.40 per second, because per-second pricing quietly assumes you keep the first take, and you won’t.

The short answer, before the detail

Yes, image-to-video works for product ads in 2026, and the economics are now boring enough to plan around. Every major lab publishes a per-second rate. A 15-second 1080p ad is a $1.80 to $6.00 generation, $6 to $24 once you count the four takes you throw away, and $14 to $42 once you count the shots, the upscaling and the subscription you do not fully use.

The thing that will actually stop an ad is not quality. It is TikTok’s product-fidelity policy, which specifically prohibits the techniques image-to-video makes easiest: faking 3D from a flat product photo, and animating a static product in unrealistic ways. Plan around that before you plan around prompts.

Pricing basis

Rates come from vendor pricing pages and API documentation checked on 16 October 2026, in US dollars unless stated. Cost examples use those published rates; they are calculations rather than measured production bills. The models were not tested for this comparison.

Why product ads are a different problem from everything else

Image-to-video is excellent at weather, crowds and camera moves. It is much weaker at hard geometry, because the model regenerates the scene every frame rather than tracking a fixed object through space. A logo mutates. A cylinder loses its curvature. A label melts into the cap by frame forty.

That is the whole game. A product ad lives or dies on whether the thing in frame is recognisably the thing you sell. So the useful question is not “which model looks best”. It is “which model holds a rigid object longest, and what does that cost per second”. Everything below is organised around that.

The 2026 model set, priced per second

Runway’s API pricing page makes the models easier to compare. It publishes a credit rate for every second of output on every model it resells, and one credit costs $0.01. It is effectively a public price list for much of the industry.

Model	Rate per second (Runway API, USD)	15-second take
Grok Imagine Video 1.5 Lite, 480p	$0.02	$0.30
Gen-4 Turbo	$0.05	$0.75
Kling 3.0 Standard, no audio	$0.09	$1.35
Veo 3.1 Fast, no audio	$0.10	$1.50
WAN 3.0, 720p	$0.10	$1.50
Seedance 2.0 Mini, 480p/720p	$0.16	$2.40
Gen-4.5	$0.12	$1.80
Kling 3.0 Pro, no audio	$0.12	$1.80
Kling 3.0 Standard, with audio	$0.13	$1.95
Veo 3.1, no audio	$0.20	$3.00
Veo 3.1, with audio	$0.40	$6.00
Seedance 2.0, 1080p	$0.40	$6.00
Seedance 2.0, 4K	$1.50	$22.50
Runway product_ad recipe, 1080p	$6.56 for a full 15s	$6.56

That table is the Runway-resold view. Google’s own Vertex AI pricing page prices Veo 3.1 at $0.40 per count with audio at 720p and 1080p, $0.20 without audio, and Veo 3.1 Lite at $0.05 at 720p with audio. Google prices in “counts” rather than seconds, but Runway, which resells Veo 3.1: charges exactly $0.40 and $0.20 per second. Matching figures on two independent vendors is strong evidence that one count is one second of output. I am flagging that as my inference, not a published statement.

Luma publishes the same kind of grid in a different unit: credits, not dollars. Useful for ranking, useless for absolute cost without the tier price in front of you.

Model at Luma	Credits per second	Credits for a 15s clip
Ray 3.14, 720p	20	300
Ray 3.14, 1080p	80	1,200
Kling 3.0, 720p, no audio	30	450
Kling 3.0, 1080p, with audio	49	735
MiniMax H3 Max, 1080p	56	840
Flux 3, FHD	102	1,530
Seedance 2.0, 1080p, with audio	126	1,890
Seedance 2.5, 1080p, with audio	191	2,865
Veo 3.1, 1080p	140 (280 with audio)	2,100–4,200

Luma’s own upscaling sits on the same page, which is handy for budgeting: upscale to 1080p is 10 credits per second, to 2K is 15, to 4K is 25.

Duration and resolution ceilings, in one place

Most 2026 image-to-video models run from 3 to 30 seconds per generation, and none of them goes past 30. For paid social that ceiling is irrelevant: your platform limits are tighter. Resolution is the number that actually constrains you.

• Kling 3.0 and Kling O3 run 3 to 15 seconds at 720p, 1080p and 4K, with optional audio and multishot prompts. Kling 3.0 4K costs $0.43 per second on Runway.
• WAN 3.0 runs 2 to 30 seconds with native audio and image, video and audio references, at 480p, 720p and 1080p.
• Seedance 2.5 runs up to 30 seconds at 480p, 720p and 1080p with a large reference budget. Its draft mode renders a 480p preview you can promote to 1080p: billed separately, at the 1080p rate.
• MiniMax H3 Max does text-to-video and image-to-video with first and last frames, at 480p and 768p.
• Luma’s API accepts 540p, 720p, 1080p and 4K, and takes frame0 and frame1 keyframes. Camera movement is controlled by language in the prompt itself; Luma exposes an endpoint listing supported camera-motion strings such as “camera orbit left”, and warns that syntactically similar phrases work “but there can be mismatches sometimes”.

Two duration notes. Seedance 2.5’s input-video surcharge means video-to-video is far more expensive per second than image-to-video: at 1080p you pay 68 credits per second of output plus 34 per second of input video. And Kling’s 15-second ceiling is a hard stop: long ads are stitched, not generated.

The four numbers the pricing pages don’t tell you

The keep rate is the real multiplier. Published per-second pricing assumes the first generation ships. In practice you will generate four takes to get one you can put in an ad. I am labelling this as a planning assumption, not a measured statistic, because I have no independent benchmark for product-shot keep rates, but it is the honest multiplier, and it is the one every cost estimate should include.

Credits per second differ between the app and the API. Runway’s consumer page says Standard’s 625 credits equal “104s of Gen-4 Turbo”, which implies 6.01 credits per second. Its API page lists Gen-4 Turbo at 5 credits per second. Gen-4.5 reconciles cleanly: 625 credits for 52 seconds is 12.02 per second, and the API lists 12. The published pages do not explain the Turbo gap; check your invoice before using that rate in a budget.

Audio roughly doubles the bill on audio-native models. Veo 3.1 is $0.20 per second silent and $0.40 with audio. Kling 3.0 Standard goes from $0.09 to $0.13. For product ads I would usually generate silent and add sound in post, unless the model is generating the sound design for you.

Upscaling is billed per frame, not per second. Runway’s Magnific video upscaler is $0.0072 per output frame at 1K, $0.0092 at 2K and $0.0123 at 4K. That is $3.24 for a 15-second 1080p clip at 30 fps. Significant, and it does not repair geometry the model already got wrong.

Runway’s product_ad recipe: cost per finished ad

Runway publishes per-second pricing for a set of named recipes, and three of them are aimed squarely at this job. The product_ad recipe costs 216 credits for a 4-second video at 1080p, plus 40 credits per additional second: 656 credits, or $6.56, for a complete 15-second ad in a single call. product_ugc is slightly cheaper at 208 credits for 4 seconds, product_swap slightly more at 228. All three allow 4 to 15 seconds per generation.

That is not comparable to a per-second model. You are buying a composite, not a clip. For a single hero ad it is the best value among the listed rates. If you need variants, you are back to per-second maths.

Cost per finished ad, worked

Here is the whole stack for a 15-second, four-shot, 1080p vertical ad. Generation uses Gen-4.5 at $0.12 per second; the keep rate is my assumption of one in four.

Line	Maths	Cost
i2v generation	4 shots × 15s × $0.12 × 4 takes	$28.80
Cloud upscale to 1080p	4 × 15s × 30fps × $0.0072	$12.96
Voiceover	~220 characters at 5 credits per 1,000	$0.01
Sound effects	15s at 1 credit per second	$0.15
Total		$41.92

Drop the cloud upscale and export at 1080p directly, or upscale on hardware you already own, and the same ad is $28.96. Now switch the four shots to Kling 3.0 Standard without audio at $0.09 per second and generation drops to $21.60, but Kling 3.0 Standard is a 720p model, so you have simply moved the cost into the upscale. That trade is the single most common budgeting mistake in this workflow.

Now the subscription case, which is where most small teams actually live. Runway’s pricing page lists Standard at $15 per month, or $12 billed annually, with 625 credits: ”52s of Gen-4.5, 104s of Gen-4 Turbo, or 78 Gen-4 Images”. At one-in-four keep rate, 52 generated seconds is about 13 usable seconds, so a 15-second finished ad costs roughly $13.85. Pro at $28 annually gives 2250 credits, or 187 seconds of Gen-4.5, which is about $8.98 per finished ad. Max at $76 annually gives 9500 credits, 791 seconds, about $5.77.

So: subscriptions win on volume and lose badly on idle capacity. Standard also caps parallel generations at five, which slows batch work considerably; Pro gives fifteen and Max twenty.

The workflow, step by step

The full pipeline takes six steps and about an hour per finished ad, most of it spent on re-rolls rather than rendering.

1. Shoot or source the conditioning still. One clean studio frame, front-lit, product square to camera, logo readable, no clutter. This sets your fidelity ceiling for everything downstream.
2. Generate or buy three to five hero stills. One clean pack shot, one lifestyle frame, one macro detail. Different angles drift very differently, and you want options before you spend on video.
3. Plan the shots before you generate. Four shots at three to four seconds each makes a 15-second vertical. Write each prompt in the five-part structure below before you open a single model.
4. Generate four takes per shot at 1080p. Not two. Then trim the first 4 to 6 frames of each clip before anything else, which kills the first-frame snap.
5. Upscale lightly and composite the logo. Cloud upsample only if the shot survives geometry checks. Overlay the real logotype on the label area in your NLE: no model holds letterforms reliably.
6. Add sound in post, then export per placement. Voiceover, sound effects and music are pennies at published rates. Export 9:16 at 1440 × 2560 for Reels and TikTok, and a separate cut for YouTube if you need one.

Prompt structures that actually hold a product still

The pattern underneath all of these is the same: shot type, camera move, subject held constant, light, what must not change. Keeping the product clause in every single prompt matters more than any clever adjective.

Luma documents camera control as prompt language, so I have written the camera moves using that vocabulary.

Turntable

Locked-off product turntable. Camera orbit left, constant speed, single complete revolution.
The matte black aluminium bottle and its white cap remain geometrically identical to the source
image in every frame. Soft top-light with a single specular highlight on the cap edge.
Do not change the label text, cap colour, bottle proportions or background colour.

Push-in

Slow push in. Camera moves forward 10 percent of frame width over the full clip, no rotation.
The amber serum bottle stays centred, unchanged in size relative to the label, unchanged in colour.
Warm side key light, black seamless background. Do not alter the dropper, the embossed logo or
the liquid level.

Lifestyle

Hand sets the ceramic mug down on an oak table and releases it. Handheld at rest, minimal movement.
The mug's glaze, handle geometry and navy band match the source frame exactly.
Warm afternoon window light from camera left. Do not change the handle shape or add steam.

Macro

Extreme macro across the brushed metal dial, shallow depth of field, focus racking from the knurled
edge to the centre mark. The engraved numerals and the indicator line remain crisp and unchanged.
Single large softbox reflection. Do not deform the numerals or move the indicator.

Slow-motion pour

Slow-motion pour into a clear glass, 120 fps look. Liquid column continuous and translucent.
The bottle remains rigid and unchanged; label, cap and glass geometry match the source frame.
Backlit against warm grey. Do not add bubbles, foam, spills or extra bottles.

Liquid

Milk swirling into black coffee in slow motion from above, marble pattern expanding outward.
The ceramic cup rim stays perfectly circular and its colour does not change.
Soft top-down light. Do not alter the cup shape, the saucer or the liquid volume outside the cup.

Fabric

Slow air movement lifting the edge of a linen shirt, fabric rippling gently toward camera.
Weave texture, seam positions and button spacing stay fixed to the source image.
Soft north-window light. Do not change the weave scale or add buttons.

Two rules that survive contact with any model. Cap your verbs: “subtle”, “gentle”, “small” beat “dramatic”. And if the product is the hero, the camera moves and the product does not. Motion belongs to the camera, the light, the liquid and the fabric: never to the object you are selling.

Failure modes and the re-roll that fixes each

Product warping and label drift. The model is regenerating geometry. Re-roll with the camera locked off and the prompt’s “do not change” clause promoted to the first sentence after the camera line. If it persists after three attempts, the conditioning frame is the problem: go back to the still.

Logo mutation. Letterforms are the hardest thing in the frame to hold. Generate the motion with a clean label plate, then composite the real logo over the label area in post. This is standard practice, not a workaround.

First-frame snap. The opening second lurches because the model reconciles the prompt against your still. Trim the first 4 to 6 frames in your NLE before you do anything else. Fixes a startling share of “cheap-looking” complaints.

Specular sparkle and flicker on reflective surfaces. Metal and glossy packaging flicker frame to frame. Kill it by lowering the light count in the prompt to one source and adding “constant exposure” as an explicit term. Failing that, apply a light temporal denoise.

Motion smear on fast moves. Any large camera move in the first two seconds smears. Start static, hold for half a second, then move. It also improves the first-frame snap at the same time.

Everything soft. You generated at 720p and are asking an upscaler to invent detail it does not have. Generate at 1080p and upscale lightly instead.

Ad platform specs you can act on

Everything below is vertical 9:16. Meta’s guidance is the strictest at 1440 × 2560 with H.264 and progressive scan; TikTok’s published floor is the loosest at just over 540 × 960. Length is the only dimension where the platforms genuinely disagree: TikTok caps at 60 seconds, YouTube Shorts has no maximum but recommends under 60, and Meta’s Reels guidance technically allows up to 15 minutes.

Platform	Format	Resolution / ratio	Length	Source
Instagram Reels ads	MP4, MOV	1440 × 2560, 9:16, H.264, square pixels, progressive, AAC stereo 128 kbps+	Up to 15 minutes	Meta ads guide
TikTok	MP4, MPEG, 3GP, AVI, MOV	9:16, above 540 × 960	5–60s, 15s recommended	TikTok Ads specifications
YouTube Shorts ads	:	Vertical	No maximum, under 60s recommended	Google Ads Help
YouTube non-skippable in-stream	(	)	15–60s by campaign subtype	Google Ads Help
YouTube bumper	(	)	5–6s	Google Ads Help

Meta’s guidance also says video sound is “optional, but strongly recommended” and captions are “optional, but recommended”. On YouTube Shorts, a TrueView view is counted at 10 seconds or the end of the video, whichever comes first, or on a CTA, like, comment or share click. A viewable impression needs 50% of the ad on screen for two continuous seconds.

Confirm the TikTok figures in Ads Manager before publishing; they have not been verified directly against the official specifications page here. Pinterest and Snapchat are omitted because their current official specifications remain unconfirmed.

Meta’s technical recommendations in the table cover Reels. Stories, feed and in-stream placements need separate aspect-ratio and safe-zone checks.

Sound, voice and music: the cheap part

Audio is where budgets leak. On Runway’s published rates it should not be. Eleven v4 costs 5 credits per 1,000 characters (about one cent for a 15-second read) and the sound-effects endpoint runs at 1 credit per second, so $0.15 for a whole ad. Note the Eleven v4 rate changed on 12 October 2026, up from an introductory 2.2 credits per 1,000 characters. Luma’s grid is comparable in spirit: 21 credits per 1,000 characters for ElevenLabs v3 text-to-speech, 25 credits per minute for sound effects, 98 credits per minute for music.

The real decision is whether you buy native audio with the video. On Veo 3.1 and Kling 3.0 that doubles your per-second cost. On a product ad, I generate silent and build the sound in post, because it gives you the licence and the mix control.

Disclosure, likeness and platform policy

EU AI Act Article 50 has applied since 2 August 2026. Deployers of systems that generate or manipulate image, audio or video constituting a “deep fake” must disclose that the content has been artificially generated or manipulated, at the latest on first exposure. The Commission’s FAQ is explicit that you cannot satisfy this by relying on the machine-readable provenance mark the model provider embedded: deployers “cannot simply rely” on it. You need a visible, perceivable disclosure. Fines reach €15 million or 3% of total worldwide turnover. Content generated before 2 August 2026 is not labelled retroactively, and there is a limited grace period to 2 December 2026 for the machine-readable marking obligation on systems already on the market.

TikTok’s AIGC policy is the one that will get product ads rejected. TikTok requires proactive disclosure when content is fully generated or significantly edited by AI, and offers both an on-screen note and an “AI-generated content” toggle that produces a “Disclosed by creator as AI-generated” label. Enabling it does not demote your content. False disclosure (in either direction) is a violation.

The product-specific clause is the part almost every competitor article misses:

“Do not use AI to turn flat (2D) products into fake 3D or rotating visuals.” “Animate static products in unrealistic ways.” : TikTok Seller University, AI-Generated Content Restrictions and Requirements

Read that next to what this article teaches you to build and the tension is obvious. A turntable shot on a flat product image is, on its face, the prohibited technique. TikTok also prohibits changing a product’s size, colour, material, quantity or bundle contents, adding lighting or glow the real product does not have, and compressing time in before-and-after visuals. Enforcement ranges from restricting distribution and removing product links, through Creator Health Rating deductions, to permanent suspension.

Meta. Meta’s public position, from April 2024 and updated 12 September 2024, is an “AI info” label triggered by industry-shared signals such as C2PA provenance data or by advertiser self-disclosure, with the label moved into the post menu for lightly AI-edited content. This is a September 2024 policy reference, not a current-policy confirmation. Check Meta’s current advertising standards before relying on it.

Likeness and rights. If any human face, voice or identifiable person appears, US FTC guidance requires disclosure of the material connection, placed in the video itself rather than only the description, with concrete language: FTC staff specifically warn against vague shorthand like “sp”, “spon” or “collab”. You cannot endorse a product you have not used, and you cannot make a claim the advertiser cannot substantiate. TikTok separately prohibits using AI to replicate anyone’s identity, likeness or voice without authorisation, and prohibits AI-generated fake professional identities endorsing health products.

Commercial licence tiers. Runway grants commercial rights and ownership of outputs on paid plans. Topaz is the one that catches agencies out: personal plans carry “personal and commercial use for orgs under $1M USD annual revenue”, and full commercial use requires Pro or Studio Pro. Above that revenue threshold, you need the pro licence.

This is not legal advice. If you are advertising in regulated categories or in multiple jurisdictions, talk to counsel.

Bottom line

Start on Gen-4.5 at 1080p, or Kling 3.0 Pro if you want Runway’s multishot prompts and native audio. Lock the camera, cap your verbs, and budget four takes per shot: roughly $30 for a finished 15-second ad, or about $14 if you commit to a subscription and actually use it.

Read TikTok’s AIGC policy before you build a single turntable, because the technique is named in it. And get the hero still right: it sets your ceiling on everything downstream.

FAQ

What does one AI product ad actually cost in 2026?

Roughly $14 to $42 for a 15-second vertical ad at 1080p. The low end assumes a Runway consumer subscription you actually exhaust; the high end assumes pay-per-second API usage with four generations per usable take plus cloud upscaling. Per-second list prices look far cheaper: $0.05 to $0.40 per second, because almost no take is publishable first time.

Which image-to-video model is cheapest per second?

On Runway’s published API table, Grok Imagine Video 1.5 Lite at 480p is $0.02 per second and Veo 3.1 Fast without audio is $0.10 per second. At 1080p the cheapest sensible tiers are WAN 3.0 and Veo 3.1 Fast at $0.20 per second, and Kling 3.0 Pro without audio at $0.12. Quality at those tiers is not comparable to Seedance 2.0 at 1080p, which costs $0.40.

Does image-to-video warp the product?

Yes, and it is the dominant failure mode. Models regenerate fine geometry every frame rather than moving a fixed object, so logos mutate, cylinder curvature drifts and labels melt. The reliable fix is a smaller motion budget, a locked-off or very slow camera, and prompt lines that explicitly freeze the product. Even then, budget for roughly one usable take in four.

Do I have to label an AI-generated product ad?

In the EU, yes. Article 50 of the EU AI Act has applied since 2 August 2026 and requires deployers to disclose artificially generated or manipulated image, audio and video that constitutes a deep fake, at the latest on first exposure. Fines reach €15 million or 3% of worldwide turnover. TikTok also requires proactive disclosure for fully or significantly AI-generated commercial content.

Can I use AI-generated product video commercially?

Usually, but read the tier, not the headline. Runway grants commercial rights and ownership of outputs on paid plans. Topaz is different: personal plans cover only organisations under $1 million in annual revenue, and full commercial use requires the Pro or Studio Pro licence. If you are an agency or an above-threshold brand, check the licence tier before delivery.

Should I generate at 720p and upscale, or generate at 1080p?

Generate at 1080p if you can afford it, then upscale lightly. Generating at 720p and upscaling four times amplifies exactly the artifacts that make product ads look synthetic: warped edges and melted type. Runway’s Magnific video upscaler is billed per output frame, so the maths are simple, but upscaling does not repair geometry the model already got wrong.

What are the Meta, TikTok and YouTube video ad specs?

Meta’s Reels ads guidance specifies MP4 or MOV, 9:16 ratio, H.264 with square pixels, fixed frame rate, progressive scan, stereo AAC at 128 kbps or higher, and a recommended resolution of 1440 by 2560 pixels. TikTok publishes 9:16 at more than 540 by 960 pixels with duration from 5 to 60 seconds. YouTube recommends under 60 seconds for Shorts ads and 15 to 60 seconds for non-skippable in-stream ads.

Is a subscription cheaper than the API?

Only if you exhaust the credits. Runway Standard at $12 per month billed annually yields 625 credits, which the pricing page equates to 52 seconds of Gen-4.5. At a one-in-four keep rate that is about 13 seconds of usable footage per month: roughly $13.85 per finished 15-second ad. Pro at $28 buys 187 seconds. If you do not use them, API pay-per-second is cheaper.

Why is my product video flagged as ‘product not as described’?

TikTok’s AIGC policy treats AI that changes a product’s size, colour, material, quantity or bundle contents as misleading advertising, and it names specific prohibited techniques: turning flat products into fake 3D or rotating visuals, animating static products in unrealistic ways, and adding lighting or glow the real product does not have.

What is the single biggest mistake people make with image-to-video product ads?

Starting from whatever product photo they already have. The conditioning frame sets your ceiling on product fidelity. A cluttered marketplace shot with a logo at an angle will drift; a clean studio frame, front-lit, with the logo square to camera, will hold far more often.

---

Sources

Pricing and model documentation

• Runway: pricing
• Runway Dev: API pricing and costs
• Runway Dev: available AI models
• Luma: plans and pricing
• Luma: video generation API docs
• Google Cloud: Agent Platform (Vertex AI) generative AI pricing
• Topaz Labs: product pricing
• Higgsfield: pricing plans

Platform policy and regulation

• TikTok Seller University: AI-Generated Content Restrictions and Requirements
• European Commission: Transparency obligations under Article 50 of the AI Act
• EU Artificial Intelligence Act: Article 50 text
• Meta: Our Approach to Labeling AI-Generated Content and Manipulated Media (April 2024, updated September 2024)
• FTC: Disclosures 101 for Social Media Influencers
• C2PA: specifications index

Ad platform specifications

• Meta ads guide: video ad specs on Instagram Reels
• Google Ads Help: About video ad formats
• Google Ads Help: ad formats, sizes and best practices