Before and after August 7 are two different corpora: 2D animation up 2.6x, prompts 36% longer
The previous piece showed that the cases cover only 47 days. This one asks whether, inside those 47 days, the content itself changed.
We split the 980 timestamped cases at the median publication time (2026-08-07 05:15 UTC) into 489 and 491, then compared every dimension.
Where the data comes from
Times are decoded from X post IDs (snowflake high bits carry a millisecond timestamp). The split uses the median, so the two halves are near-equal by construction. Style and technique tags remain automatic candidates plus human review; unreviewed items are excluded.
Finding 1: 2D animation goes from 9.2% to 24.4%
This is the largest movement of any dimension:
| Visual style | Earlier half | Later half | Change |
|---|---|---|---|
| 2D animation | 9.2% | 24.4% | +15.2pp |
| Cinematic | 47.2% | 55.2% | +8.0pp |
| 3D animation | 6.3% | 11.8% | +5.5pp |
| Black and white | 1.4% | 2.6% | +1.2pp |
| Photoreal | 44.2% | 45.6% | +1.4pp |
| Slow motion | 14.9% | 14.9% | 0.0pp |
2D animation grows 2.6x in the later period. Our reading: early on, people tested whether they could make something that looks like live action — the most legible proof of model capability. Once that was established, attention moved toward animation, a style that is safer for the model (no photoreal face to get wrong).
Finding 2: Later cases are markedly more multi-shot
| Technique | Earlier | Later | Change |
|---|---|---|---|
| Cut timing | 13.7% | 23.6% | +9.9pp |
| Multi-shot switching | 11.7% | 20.8% | +9.1pp |
| Background score | 8.0% | 16.9% | +8.9pp |
| Environment foley | 4.1% | 12.4% | +8.3pp (3x) |
| Prop interaction | 12.9% | 17.9% | +5.0pp |
| Dolly | 14.5% | 20.2% | +5.6pp |
| Emotional performance | 22.9% | 32.6% | +9.7pp |
| Lip-sync | 22.9% | 30.5% | +7.6pp |
| On-screen text | 22.1% | 27.5% | +5.4pp |
| Weapon state | 13.3% | 9.8% | −3.5pp |
| Reference image | 20.2% | 16.7% | −3.5pp |
That last group is the only one going down. Cut timing and multi-shot switching rise while weapon state falls — the signal of a shift from in-shot effects work toward cutting between shots.
Environment foley tripling is a strong signal too: early cases had picture and score; later ones started handling sound detail inside the shot.
Reference images dropping 3.5pp says later work is more often generated from scratch rather than replicated from a reference.
Finding 3: Prompts grew 36% longer
| Metric | Earlier half | Later half |
|---|---|---|
| Median prompt length | 1,498 chars | 2,035 chars |
+35.8%. This is the same fact as "multi-shot switching +9.1pp" seen from another angle: writing multiple shots requires describing camera, action and transition for each one.
Finding 4: Genres slide from demo to commercial
| Genre | Earlier | Later | Change |
|---|---|---|---|
| Fashion & beauty | 23 | 56 | +143% |
| Transition editing | 9 | 22 | +144% |
| Character performance | 53 | 75 | +42% |
| Tool comparison demo | 15 | 26 | +73% |
| City & architecture | 35 | 44 | +26% |
| Short-drama dialogue | 75 | 80 | +7% |
| Product ad | 80 | 63 | −21% |
| Talking-head UGC | 21 | 6 | −71% |
| Action & physics | 23 | 12 | −48% |
| Music & dance | 47 | 35 | −26% |
Talking-head UGC falls 71% and action & physics 48%, while fashion & beauty and transition editing more than double.
The two early leaders (product ads, talking heads) are both demonstration content — proof that the tool can do the job. What grew later is aesthetic content (fashion, character performance) and technical content (transition editing).
One related framing shift: vertical clips go from 59 to 84 (+42%), consistent with fashion & beauty's +143% — that is the most vertical genre.
What you can do with this
If you learn prompt craft from public cases: note when the case you are studying was posted. An early-August and an early-September case are structurally different objects — the earlier one is roughly 500 characters shorter and usually describes a single shot.
If you build tools: multi-shot switching and cut timing rising sharply means shot-level management is becoming a real need. That is also why "first/last frame" tooling appeared later.
If you pick topics: the early favourites (talking heads, action) are receding while fashion & beauty and transition editing climb. This is not a judgement about what content works — only about where public attention sits.
Method and limitations
- Sample of 980 (those with resolvable times), not the full 1,274. GitHub and tutorial sources lack timestamps and typically lag, which may understate the trends here.
- The split point is the median, so this is "relatively earlier vs later" rather than natural periods. A different split would move the magnitudes (direction should hold).
- Tag scope unchanged: style and technique remain automatic candidates plus human review with unreviewed items excluded. So absolute levels for labels like "2D animation" depend on review progress; this piece compares relative change only.
- No causal claim: this reports co-occurring change. "Animation grew because models improved" is our interpretation, not a data finding.
- All figures use 980 of the 1,274 index records.
Reproduce it
Each record carries a publication date decoded from its source link and can be browsed chronologically: https://cases.aishifu.shop/