Data window: 13 July – 18 October 2026. All figures come from the site's own prompt log; nothing here is a general benchmark.
Most "prompt guides" are written from a handful of hand-picked examples. We run an image and video generation site, and every request that reaches our workers is logged. This report summarises that log: which models actually got used, how long our prompts are, which photographic phrases show up most often, and, just as important, how little of it you should read as advice.
How the data was collected
- Source: the `user_prompts` table of the ZView Space database (one row per submitted generation request).
- Window: 13 July 2026 to 18 October 2026 — 25,253 requests.
- Who submitted them: 59% (14,887) were submitted from the operator's own admin console, which we use to produce short-form vertical video content. 29% (7,286) came through the public web tools. The remaining 12% came from image-to-video, image-to-image and video pipelines.
- What we did not include: no individual prompt text, no user identifiers, no gallery images. Everything below is an aggregate count or a length percentile.
Because most requests are operator-driven, this is a picture of one production workflow plus a small public audience, not of the average AI-image user. We say that up front because it changes how every number below should be read.
Which models did the work
| Model | Requests | Share | | --- | ---: | ---: | | Krea2 Turbo (image) | 21,110 | 83.6% | | MiniMax H3 (video) | 1,730 | 6.9% | | Z-Image Turbo (image) | 1,335 | 5.3% | | Wan (video) | 670 | 2.7% | | Anima (image) | 408 | 1.6% |
Krea2 Turbo carries the workload because our vertical-video pipeline starts from Krea2 stills. Z-Image Turbo, by contrast, is mostly used through the public tool: 1,240 of its 1,335 requests came from the web form and none from the admin console.
Monthly volume peaked in August (11,197 requests) and came down to 4,279 in September and 2,951 in October so far (to the 18th).
Format: almost everything is vertical
91.5% of all requests (23,098) were 9:16. Only 5.5% were 16:9, 2.5% were 4:5 and 0.5% were 1:1. If you are choosing a default aspect ratio for a tool built around phone-first video, the log says vertical should be the default, not an option buried in a menu.
How long are the prompts?
Median prompt length in characters (10th and 90th percentile in brackets):
| Model | Median | P10 – P90 | | --- | ---: | ---: | | Krea2 Turbo | 1,094 | 384 – 1,771 | | Anima | 887 | 483 – 1,626 | | Z-Image Turbo | 563 | 109 – 1,548 |
Two things stand out. Krea2 prompts are roughly twice as long as Z-Image prompts at the median, which matches how we use them: Krea2 prompts are templated and include camera, light and framing blocks, while many Z-Image prompts are a short sentence typed into the public form. And the spread is wide, so "ideal prompt length" is not a single number — it depends on whether you are writing a one-line idea or a structured shot description.
Of the Krea2 prompts, 91.3% are unique (19,277 distinct prompts out of 21,110), so the long length is not just the same template resubmitted.
Which photographic phrases appear most
We counted how many Krea2 prompts contain each phrase (case-insensitive substring match, so "pore" also catches "pores"):
| Phrase | Prompts | Share of Krea2 prompts | | --- | ---: | ---: | | candid | 9,915 | 47.0% | | 85mm | 8,316 | 39.4% | | pore | 8,218 | 38.9% | | skin texture | 5,350 | 25.3% | | cinematic | 4,217 | 20.0% | | studio | 2,944 | 13.9% | | shallow depth | 2,601 | 12.3% | | close-up | 2,130 | 10.1% | | street | 2,089 | 9.9% | | editorial | 1,469 | 7.0% | | flash | 1,297 | 6.1% | | bokeh | 728 | 3.4% | | golden hour | 650 | 3.1% | | film grain | 627 | 3.0% | | 35mm | 343 | 1.6% | | f/1.4 | 321 | 1.5% |
Read this table carefully. "85mm" and "pore/skin texture" are common because our own templates contain them, not because users independently converged on them. What the table does show is the vocabulary that a realistic-portrait workflow settles on when you iterate on it for months: a focal length, a natural-skin cue, a light or setting word, and a framing word. Phrases that sound impressive but barely appear (for example "backlit", 11 prompts) are worth noticing too, because we rarely needed them.
What this log cannot tell you
- It does not say which phrases make better images. We did not score output quality, and the like counts on our gallery are too small (a few hundred in total) to say anything statistically.
- It is not a sample of "users". Most requests are from one operator running a repeatable pipeline.
- Public usage is small and uneven. Public web requests were concentrated in July and August; the figures above should not be used to estimate market demand.
Takeaways we actually act on
1. Default new tools to 9:16, and keep the other ratios one tap away. 2. For Krea2, write prompts as structured blocks (subject, setting, light, camera) and expect to land around 1,000 characters; for Z-Image, a short prompt is normal and fine. 3. Keep a small shared vocabulary for realism cues rather than inventing new phrases each time; the repeated phrases above are what keep a batch visually consistent. 4. Log what you generate. Without the log, none of the numbers above would exist.
Method note: counts are produced by read-only SQL aggregation against the production database on 19 October 2026. Percentiles are computed over character length of the stored prompt text, including very long outliers, which is why P90 rather than the maximum is reported.