Ad variations: how to make many versions of one video ad
Ad variations are versions of one ad that differ in one element. Plan the matrix, make each version once, check stills first, and count the cost.

Ad variations are versions of one ad that differ in a single element, such as the hook, the presenter, the message, the product shot, the call to action, or the frame shape, so a test can show which change made the difference. Make them by holding everything else fixed. The number of versions is the product of the options you vary, and so is the cost.
With Sume, each version is its own job or Format run with its own idempotency key, and stills can be checked before the full render. The facts come from Generate avatar video, Avatar video previews, Jobs and results and Bulk runs, read on 2026-09-28.
Which elements of a video ad can I vary?
Split the ad into parts you can swap one at a time. On a Sume avatar video, a multi-scene video_inputs list gives each part its own scene, like the docs' example with scenes named hook, demo and cta:
| Element | What changes | Where it lives on Sume |
|---|---|---|
| Hook | The opening line and shot | The first video_inputs scene; or one Timeline render per hook, see hook variations |
| Presenter | Who speaks | avatar_handle; one avatar per final video |
| Message | The body copy | The middle scenes' spoken text |
| Product shot | How the product appears | product_image, or the scene prompt or photo |
| Call to action | The closing line | The last scene |
| Frame | The shape | aspect_ratio: 1:1, 3:4, 9:16, 4:3, or 16:9 |
How many variations will I end up with?
Multiply the options. 3 hooks × 2 presenters × 2 calls to action is 12 versions. If you need to know which change mattered, vary one element per set and keep the others at a fixed control; crossing every element makes the count, and the bill, grow fast.
How do I make each version once and only once?
Give every version its own Idempotency-Key, built from the variant itself (for example sku-1042-hook2-presenterB-v1), not from the time of asking. A retry with the same key and the same body returns the original job instead of billing a second one, so a retry after a dropped connection does not make a version twice. On a Format run, the same key with a different body is refused with 409 idempotency_conflict. Bump the v1 only when you really want a new take.
Can I check variations before paying for full renders?
Yes, for avatar videos. POST /v1/avatar-video-previews takes the same fields as the video and makes only the first-frame stills, one per scene for multi-scene video_inputs, so you can approve composition before spending a full render. POST /v1/avatar-video-previews/{id}/regenerate refreshes the stills, and POST /v1/avatar-video-previews/{id}/generate-video starts the real video from an approved preview. Changing the script, scenes, avatar or aspect ratio needs a new preview. More in avatar video first-frame previews.
How do I run all the variations at once?
If your versions are Format runs, one bulk request can queue them: items holds 1–100 run bodies, one per version, and concurrency (1–16) sets how many run at once. Give each item the variant's own brief, so the queue is your matrix written out. How the queue works, and how to track it, is in Sume Format bulk runs.
How much do ad variations cost?
Cost scales with versions × length. At the plus tier with a product image, $0.258 per second, the 12 versions above at 20 seconds each come to $61.92, plus a 5.5% agent fee by default. On Format runs, generation_spend_cap_usd caps each version's generation, up to $500 per run. The full rate card is on API pricing.
Sources
Related posts
More in Use cases
- Product shadow: add a natural shadow to a product photo
A product shadow grounds the item on its surface. Cut-outs lose it; add it back with an AI image edit, or keep it by changing only the background.
- How to add a video to a Shopify product page
Add a video to a Shopify product page from the product's Media section: upload the file or paste a YouTube or Vimeo link. Your theme must support video.
- AI 360 rotation video: spin a product from two photos
From one photo, AI invents every side it can't see. For a 360 rotation video, pin the front and back as frames, make two half turns, and join them.
- AI avatar for events: a host for screens and lobby loops
An AI avatar can host an event's scripted parts: the welcome, housekeeping, sponsor thanks, and session intros as short 16:9 clips, joined or looped.
Written by Sume