90% of classifiable AI videos use one of just two visual styles

By Alpha Lay · 2026-09-26

We defined 8 values for visual style: cinematic, photoreal, 2D animation, 3D animation, slow motion, stop-motion collage, black and white, plus "unreviewed."

After the statistics ran, the share of the top two values came out high enough that we re-checked the code. It wasn't a bug.

Where the data comes from

Style tags are automatic candidates plus human review. 421 cases (33.0%) carry the "unreviewed" marker, making this the slowest dimension to review of the four. To keep the conclusion free of missing-value contamination, every ratio below uses the 853 cases carrying at least one reviewed style tag as the denominator.

Finding 1: 767 of 853 cases use one of only two answers

The top two values sum to 1,155 — more than the case count, because a case can carry both tags. So we de-duplicated:

90% of classifiable cases land on two visual answers.

Broken out further:

The largest group carries both (388) — in most cases these two are two descriptions of one thing: shooting real people so they look real, with cinematic lighting and depth of field.

Finding 2: What the other 86 look like

The 10.1% that use neither answer have a clear composition:

Those 86 are essentially the whole "animation / handmade texture" family. Which is to say the aesthetic map of the public corpus reduces to a stark binary: either it pursues looking like filmed cinema, or it is explicitly non-photoreal animation. The middle ground — illustration styles, picture-book, pixel art, collage experiments — barely exists across 1,274 cases.

Finding 3: Black and white appears 24 times, and not as a flex

We looked at the lowest value separately:

Not one of the 24 is a martial-arts clip. The bulk sits in music & dance and character performance — the categories that depend most on mood and least on information. In this corpus, black and white is not a stylistic experiment. It is a mood tool.

What you can do with this

If you are setting a visual direction for a new project: knowing which two answers 90% of people chose is itself decision input. Picking "cinematic" puts you in an extremely crowded lane where you must differentiate through content and hooks. Picking outside the two — stop-motion collage, at 66 cases — gives you very little reference but inherently higher recognisability.

If you wonder why your output "looks like everyone else's": the cause may not be the details of your prompt but the level of style selection itself. A 90% convergence means the space of style vocabulary has already been compressed very small.

If you evaluate video models: almost all public benchmarking currently happens inside the "cinematic plus photoreal" combination. Judging an animation model by that aesthetic will systematically underrate it.

Method and limitations

• The 421 unreviewed cases are the biggest risk: 33% of samples are unreviewed. If their style composition differs from the reviewed set, these ratios will move. We cannot currently verify this.
• Tags are "candidates plus review", not strict double annotation. Different annotators may place the "cinematic" threshold differently.
• "Cinematic" is not defined precisely in the schema — a real practical ambiguity we list as a to-improve item.
• All figures use the 1,274 index records; after de-duplication by original publication (924 cases) the convergence conclusion points the same way.

Reproduce it

Filter by visual style, optionally crossed with genre: https://cases.aishifu.shop/