aishifu Research H3 Case Library About Contact 中文

73% of AI video cases don't say how they were made — and the reason isn't the creators

By Alpha Lay ·

When we designed the tag schema we included one basic field: how was this video generated — text-to-video, image-to-video, reference-to-video, or first-and-last-frame?

It is the precondition for learning anything from a prompt. To study how a case was written, the first thing you need to know is whether it had a reference image or a specified first/last frame, because those three prompt styles are written completely differently. After filling in all 1,274 records, more than seventy percent of that field is empty.

Where the data comes from

The generation-mode field was determined from three sources: explicit statements on the original publication page, structural features of the prompt text itself (for example, whether it declares a first frame), and official labels on source sites. When none of the three yields an answer, we record "mode unverified" rather than guessing.

Generation modeCasesShare
Mode unverified93073.0%
Text-to-video16412.9%
Reference-to-video12810.0%
First/last-frame-to-video493.8%
Image-to-video (unsplit)30.2%

Our first hypothesis was that creators are simply lazy about writing it down. Slicing by genre appeared to support that — the spread between highest and lowest is nearly 50 percentage points:

GenreUnverified rateSample
Action & physics94.3%70
Transition editing86.0%43
Short-drama dialogue79.0%224
…
Sci-fi53.6%56
Martial arts / wuxia45.1%71

It looks like different communities have different habits. But the real cause is something else.

Finding 1: The missing rate is a mirror of the source channel

We re-cut the data by provenance — where each case came from:

ProvenanceSampleUnverified rate
official60.0%
source_page_public_prompt6468.8%
creator-verbatim (copied from the original post)64552.4%
beatapi_indexed_prompt_with_webm312100.0%
indexed_from_awesome_minimax_h3_prompts74100.0%
apimodels_original_or_indexed63100.0%
official_or_official_index_prompt37100.0%
Republished by static.minimax-h3.io20100.0%

The pattern is perfectly clean: every official primary channel has a 0% unverified rate; every third-party aggregation mirror has a 100% unverified rate.

That is a far stronger result than the genre cut. It says "mode unverified" is not a property of creators — it is information stripped out during propagation. The original post often says it. The aggregation site scrapes it, that line disappears, and every downstream indexer is permanently in the dark.

Finding 2: The only two complete sources aren't the community's main channels

Aligning hosting providers against unverified rate makes it starker:

Video hostSampleUnverified rate
pub-…r2.dev (official republishing channel)200.0%
h3-field-notes-production.up.railway.app64852.2%
video.twimg.com (native X video)15592.9%
media.beatapi.io312100.0%
raw.githubusercontent.com60100.0%
external-cdn.morphic.com44100.0%
static.minimax-h3.io20100.0%

Meanwhile 87.4% of the whole corpus originates on X (Twitter) — which happens to be the venue least likely to carry structured metadata.

Finding 3: Among the 344 verified cases, short drama under-indexes

Restricting to the 344 cases with a verified mode, the genre mix differs from the whole:

GenreVerified cases
Product ad57
Character performance51
Short-drama dialogue47
Martial arts39
City & architecture34

Short-drama dialogue is the largest genre overall (224 cases, 17.6%) but ranks only third in the subset that can say how it was made (47 cases, 13.7%). The more a genre depends on specific reference material, the more likely its generation mode is lost in propagation.

What you can do with this

If you learn prompt craft from public cases: verify the generation mode first. If you cannot, don't copy the prompt. The same line — "the camera slowly pushes in" — is written very differently for text-to-video versus reference-to-video, and copying across layers wastes hours. Our advice: only study cases with a verified mode. In this library there are 344 of them.

If you build datasets or indexes: this is worth noting down. Scraping from aggregation sites systematically drops structured fields, and the loss is not random — the more thorough the aggregation, the cleaner the fields get. To preserve this class of information you have to capture it further upstream.

If you publish your own work: state the generation mode in the post body, even just "reference-to-video." It costs nothing and materially raises the chance your case is classified correctly by downstream indexers.

Method and limitations

Reproduce it

Filter by generation mode, or group by source site: https://cases.aishifu.shop/