Only 15% of 1,274 AI videos are vertical — and short-drama clips rank 16th of 18

By Alpha Lay · 2026-09-26

For the past few weeks we have been doing something tedious: indexing 1,274 publicly shared H3-generated videos one at a time, tagging each with genre, shot technique, visual style, aspect ratio, duration and generation mode.

One result we did not expect: at a moment when every short-form platform is vertical-first, only 15% of these public examples are vertical — and short-drama clips are among the least vertical of any genre.

This piece pulls out the aspect-ratio numbers on their own, because it is the dimension with the sharpest contrast and the most direct bearing on day-to-day work.

Where the data comes from

The 1,274 cases were collected from public sources, and we kept the original publication link for every one. Aspect ratio was read from the video file itself, not inferred from the poster's description. Genre and technique tags were generated as automatic candidates and then reviewed by a human; anything not yet reviewed keeps an "unreviewed" marker and is excluded from the genre statistics below.

Finding 1: Landscape leads 5.6 to 1

Vertical plus square comes to 214 cases, or 16.8%. That in itself is unsurprising — AI video tooling and viewing contexts have centred on landscape for a long time. The surprising part is next.

Finding 2: The most vertical genres are not narrative ones

First, where do the vertical cases cluster? (A single video can carry more than one genre tag.)

Read this table alone and short drama sits in the top four, which feels intuitive. But switch to each genre's own vertical rate and the conclusion flips:

Short-drama dialogue has 224 cases — more than any other genre — but a vertical rate of 10.7%, ranking 16th out of the 18 genres with enough samples to judge. In other words, the short dramas in this public corpus were overwhelmingly shot landscape.

And the two most vertical genres — visual design and fashion & beauty — are both *display* content, not *narrative* content. In this dataset, vertical framing behaves more like a product-showcase format than a storytelling format.

Finding 3: Clips with speech are almost all landscape

We looked separately at the 251 cases where speech was detected:

Among clips with dialogue, landscape is *more* dominant than in the overall set. And the median duration of vertical and landscape dialogue clips is exactly the same — 15.2 seconds both. So the common assumption that "vertical clips are shorter" does not hold here.

Speech-bearing cases concentrate in short-drama dialogue (109), character performance (40) and music & dance (30).

Finding 4: Duration piles into 8–16 seconds

86% lands in a single bucket. Short-drama dialogue has a median of 15.2s and a mean of 16.2s — pinned right at the upper edge. That suggests the bucket's upper bound is the model's reliable single-generation limit, not a duration creators actively chose. (We took that "15 seconds" apart in a separate piece; it is more extreme than it looks here.)

What you can do with this

If you make vertical short drama (Douyin, TikTok, Reels): this public corpus is close to useless as direct reference, because 89% of the short-drama samples are landscape. What you need is vertical composition and vertical camera craft, and the closest public material is visual design and fashion & beauty — genres whose vertical rates are three times higher.

If you are evaluating model capability: the narrow 8–16s bucket matters more than the aspect-ratio gap. It means the practical usable unit of these models is roughly a dozen seconds; anything longer has to be built by chaining segments, not by asking for more in one go.

If you make product ads: a 15.6% vertical rate is in line with the overall figure, meaning this genre has no strong framing preference. Decide by placement channel.

Method and limitations

• Sampling bias: only publicly released cases enter the statistics, skewing toward better-looking work. Not representative of all creators.
• Tag provenance: genre and technique tags are automatic candidates plus human review. 1,212 records are still marked "searchable, pending semantic review"; unreviewed markers are excluded from the genre statistics.
• Aspect ratio and duration come from the video files themselves — objective measurements, with error limited to container metadata accuracy.
• Deduplication scope: the ratios above use all 1,274 index records. The cross-host duplication problem is covered in a separate piece; after de-duplication the aspect split (82.5% landscape / 15.3% vertical / 2.3% square) points the same way.
• Correlation is not causation: this piece reports distributions only. "Short drama is landscape because the publishing venue is Twitter" is a hypothesis, not a finding.

Reproduce it

All 1,274 cases — with original links, aspect ratio, duration and tags — can be filtered and checked case by case at https://cases.aishifu.shop/. Every number here can be reproduced with the filters there.