3 days free on every plan · Try any plan before you pay.

Start free trialStart

BlogCreative testing

AI video variations: turning one ad into hundreds of versions

What makes a video variation read as a new video, which edits don't count, and what 31,973 posts of versioned videos show about why volume matters.

By Abondy ResearchPublished Updated 6 min read

In short: AI video variations are versions of one video, each changed enough to post as a new video rather than a copy. In our own look-alike tests, a mirror, a colour filter, a small zoom, a border, a slight turn or reordered parts each still read as the original. The sound was the giveaway until each version got its own speed and pitch. Volume is the point: in our data, the median upload's typical post got about 59 views and the median best post about 2,100, and 84.2% of uploads had a post past 1,000 views within their first 20 posts. From August 29 to November 12, 2026, our video model Flagship made 17,920 versions from 858 uploads, about 21 per upload.

This post covers what a variation is, which changes make it a new video, how many you need, and what the versions did once real people posted them. We make Flagship 1.2, so the specifics come from how it works, and you can see its versions on Explore.

What is an AI video variation?

A variation is the same video with a different look: a new crop, layout, caption or sound. It isn't a new creative. A new creative changes the idea, such as a different hook, script or product shot. Variations of one creative let you post the same idea many times, on many accounts and apps, without each post being the same file.

That matters for two reasons. Apps can treat repeated uploads of one video as duplicates; we cover that in posting the same video on multiple TikTok accounts. And views per post vary so much that one post of a video tells you little about it, so you need many posts to see what a video can do.

Which edits make a video read as new?

We tested common edits with our own look-alike check on one 7-second video on October 10, 2026. This is our check, not TikTok's, YouTube's or Meta's, and we don't know exactly how each app compares videos.

Edit, on its ownDid it still read as the original?
MirrorYes
Colour filterYes
1.15× zoomYes
BorderYes
4° turnYes
Parts reorderedYes
Untouched soundYes: every copy's audio lined up with the original's
Sound 4% to 14% quicker, with a pitch changeNo: the audio match dropped from 0.87 to about 0.03

None of the quick edits worked alone; each still read as the original and as each other. The sound mattered most: left untouched, it matched the original even when the picture had changed, and it tipped borderline pictures into "copy".

So each Flagship version gets its own speed and pitch, plus one picture change big enough to count on its own:

  • a 1.25× to 1.35× crop, each from its own corner of the frame,
  • a picture-in-picture,
  • a card, or
  • a frame, such as a tweet, a YouTube description sheet, a chat or a headline post.

Mirroring, turns, colour grades and grain were dropped on October 29, 2026. They didn't move our check, and they looked like cheap edits. Each version also gets its own caption, and the set keeps the shape of the video you upload: tall, wide or square. Before posting, each version is checked against the others, and versions that look too alike are held back.

How many variations does one ad need?

More than feels natural, because the views on one video's posts spread wide. These are the 272 uploads in our data with at least 20 posts whose views we read, made by all our video models:

One upload, across its postsMedian across uploads
Posts read100
Different versions posted43.5
Views on the middle post59.25
Views on the best post2,100
Best post ÷ middle postabout 53×

The middle upload's best post had about 53 times the views of its typical post. Across uploads, the typical post ranged from about 9 views (the 10th percentile) to about 156 (the 90th).

How quickly did an upload find a post past 1,000 views?

Posts read, in the order they went upUploads with a post past 1,000
First 565.8%
First 1076.5%
First 2084.2%
All posts read (median 100)91.5%

And every one of them had a post past 100 views within the first 20. So plan on 20 or more posts on each app you care about, rather than a handful of variations posted once each. Twenty versions on four apps is 80 posts.

Do versions do the same on every app?

No. We took versions posted on two apps and compared the first post on each. The rank correlation (1 = same order, 0 = no link) was:

Same version onRank correlation
TikTok and YouTube Shorts0.074
Instagram Reels and Facebook0.177
TikTok and Facebook0.092
TikTok and Instagram Reels0.050
YouTube Shorts and Facebook0.025
YouTube Shorts and Instagram Reels−0.017

A version that did well on one app said almost nothing about how it would do on another. In practice, a version is usually posted once per app. Of the 5,037 Flagship versions with counted posts, the median went up 3 times and 70.5% went up on two apps or more. None went up more than 4 times.

Flagship's versions on their own looked like this: 136 uploads with 20 or more posts we read, a median of 60.5 posts and 23.5 versions each. The median upload's typical post had 42.75 views, and the median best post 1,533. 83.1% had a post past 1,000 views, and 4.4% one past 10,000. We didn't match campaigns or weeks between Flagship and the other models, so don't read the two tables as a contest between models.

What does it cost to make and post variations?

On an Abondy plan, making versions uses your free videos first: 100 a month on Plus, 1,000 on Pro and 2,500 on Ultra. Posting a version costs 12 credits, about 16¢ to 20¢ a post, and since November 20, 2026, you're charged only for posts that reach 100 views within 24 hours. Plans start at $20 a month. With an API key, it's $0.05 a video made and $0.30 a post, from a prepaid API balance; the API docs cover the make call, which takes 1 to 400 versions.

How we counted

  • Sample: posts made from August 29 to November 12, 2026 (New York time), pulled on November 20, 2026, when every post was at least 7 days old. 31,973 counted posts were versions made by our video models, from 395 uploads.
  • Versions made: the 17,920 Flagship versions are every version finished for jobs started in the window, posted or not.
  • Exclusions: Abondy's own test accounts, posts whose views were confirmed as bought, and posts that were flagged, rejected or deleted.
  • Updated November 21, 2026: an earlier version counted posts whose views we never managed to read as 0 views. They're now left out (11.3% of the version posts in this sample), and the view figures were recounted on November 21, 2026 from the 28,356 version posts we read. How many times each version went up still counts every post.
  • Look-alike tests: our own check on one short video, October 2026. A platform's own systems may judge differently.
  • Instagram: read the Reels numbers as a floor. Some counts come from each post's public page, and in a 16 Nov 2026 check of 37 posts that page showed about a third of the real count.
  • Associations, not proof: these numbers describe our posts, not what a given video of yours will do.

Questions people ask

What are AI video variations?

Versions of one video with a different crop, layout, caption or sound, made so each can post as a new video. They keep the same idea; a new hook or script is a new creative.

Is mirroring or a filter enough to make a video new?

Not in our tests. A mirror, colour filter, 1.15× zoom, border, 4° turn or reordered parts each still read as the original. The untouched sound was the giveaway.

How many variations should I make of one ad?

Enough for 20 or more posts per app you test on. In our data, 65.8% of uploads had a post past 1,000 views within their first 5 posts, and 84.2% within 20.

Does a variation that wins on TikTok win on Reels?

Not in our data. The same version's first posts on TikTok and Instagram Reels had a rank correlation of 0.050.

Read next