How people actually use AI video generators: 1,529 real videos (2026 data)
Opens the editor in image-to-video mode, the way most videos in this study started. Failed and refused runs are refunded automatically.
Animate a photoMost writing about AI video is about what the models can do. This post is about what people actually ask them for. It uses Viralance's own records: every video job between 29 January and 31 October 2026, with our own team, test and banned accounts removed, and restyles of existing videos left out. 60% of the jobs happened in October, so the numbers mostly describe autumn 2026.
One thing shapes almost every table below. New accounts get a few free credits for the trial models (VEO 3.1 Lite, MiniMax H3 Max Turbo), and the most common trial clip by far was VEO 3.1 Lite at 6 seconds and 720p (509 of its 916 videos). So the "all videos" columns lean towards what a first try looks like. The "paying customers" columns (34 people, 208 videos) show what people choose once they have paid. That second sample is small, so read its percentages as a direction, not a precise share.
Most videos start from a photo
| How the video started | All videos | Paying customers |
|---|---|---|
| Photo (image-to-video) | 805 (52.6%) | 136 (65.4%) |
| Text prompt only | 620 (40.5%) | 37 (17.8%) |
| Reference images (character or product) | 58 (3.8%) | 24 (11.5%) |
| Extending an earlier clip | 4 (0.3%) | 4 (1.9%) |
| Not recorded | 42 (2.7%) | 7 (3.4%) |
Animating an existing picture is the main use, and it grows once people pay: 77% of paying customers' videos started from a photo or from reference images, against 18% from text alone. People bring something they already have, such as a product shot, a listing photo or a portrait, and want it to move.
Photos are also where most failures come from, because the provider's safety check reads the picture as well as the prompt. We cover that in why image-to-video fails.
Vertical wins, but not for everyone
When the video starts from a photo, the shape follows the photo. When people type a prompt and pick a shape themselves, the answer is clear:
| Shape chosen for a text-to-video prompt | Videos |
|---|---|
| 9:16 vertical | 549 (93.4%) |
| 16:9 horizontal | 36 (6.1%) |
| Other (3:4, 21:9) | 3 (0.5%) |
588 text prompts had a shape recorded. The vertical share is partly the editor's default, but the people who changed it are telling: among paying customers' videos with a chosen shape, 22 of 60 were horizontal (37%). People making ads for TikTok, Reels and Shorts stay vertical; people making a website banner, a YouTube intro or a property tour switch to 16:9. Our guide to reframing one video for every platform covers the case where you need both.
Short clips, mostly at 720p
| Clip length | All videos | Paying customers |
|---|---|---|
| 4 seconds | 8.4% | 12.0% |
| 5 seconds | 15.2% | 29.3% |
| 6 seconds | 47.9% | 20.7% |
| 8 seconds | 17.2% | 18.3% |
| 10 seconds | 6.5% | 7.7% |
| 15 seconds | 4.4% | 10.6% |
Six seconds dominates the overall table because it is the most common VEO 3.1 Lite trial clip. Paying customers spread out: 5 seconds was their most common length (many Kling, Seedance and PixVerse clips come in 5-second steps), and they chose the longest option, 15 seconds, more than twice as often. Longer AI clips are usually made by stitching or extending; see how long can AI videos be.
Resolution followed the same pattern. 720p made 60.6% of all videos and 48.1% of paying customers' videos. 1080p was 13.7% overall and 18.3% for paying customers, and 2K or 4K was chosen 8 times in total. 1080p costs more and waits longer, which is why most people only use it for the final version. The trade-off is in 720p vs 1080p for AI video.
Paying customers mix models
| Model | Share of paying customers' videos after their first purchase |
|---|---|
| VEO 3.1 Fast | 26.7% |
| VEO 3.1 Lite | 17.0% |
| Seedance 2.0 | 11.9% |
| MiniMax H3 Max Turbo | 11.4% |
| Grok Imagine | 10.8% |
| Kling 3.0 Pro | 10.2% |
| Others (9 models) | 12.0% |
176 videos by paying customers had a known purchase time and were made after it. Across all accounts, VEO 3.1 Lite made 59.9% of videos, because it is the main free-trial model. Once people pay, no model passes 27%. The typical paying customer used 2 different models, and across all accounts 27 models were used at least once.
In practice that usually means picking a model per job: a fast, cheap one for drafts, Seedance 2.0 or Kling 3.0 Pro for a hero shot, and a model that accepts photos of people when another one refuses. A tool tied to one model works against that habit. On Viralance every model is paid from the same credit balance. Our model comparison has the differences side by side.
Paying customers also spent more per video: 9.1 credits on average, against 5.1 for videos made on free credits.
What people make
We matched prompts against keyword lists. A prompt can count for more than one topic, and keyword matching misses things, so treat this as a rough map rather than exact shares.
| Topic (keyword match) | Share of prompts | Finished |
|---|---|---|
| Camera directions (drone, dolly, close-up, slow motion, cinematic) | 48.0% | 92.0% |
| Products, brands and ads | 37.7% | 92.4% |
| People (a woman, a man, a model, a selfie, a portrait) | 37.1% | 87.8% |
| Homes and interiors | 10.8% | 93.9% |
| Landscapes and weather | 9.6% | 94.6% |
| Fantasy and sci-fi | 8.5% | 93.1% |
| Cars and motorbikes | 5.7% | 89.7% |
| Animals and pets | 5.6% | 94.1% |
| Food and drink | 4.4% | 91.2% |
| Halloween | 3.5% | 86.8% |
Two things stand out. Almost half of all prompts direct the camera, so people think in shots, not just in scenes. And commercial work (products, ads, property) is a large part of the volume, next to personal and fun videos. Prompts with people in them finished least often of the large groups, for the same reason photos do: the provider is more careful with people.
The median prompt was 68 words long (a quarter were under 30, a quarter over 139). That includes text from templates and from the prompt helper, so it says more about what works than about what people type from scratch.
Short prompts are not the problem
At first glance, very short prompts looked risky: photo jobs with a prompt of 10 words or fewer finished 60.4% of the time, against 86.4% with longer prompts. When we read the refused short prompts, most of them asked for undressing or sexual motion. The provider refused the request, not the length. With text-to-video, where that kind of request was rare, short prompts finished 96.2% of the time (25 of 26). A short, clear prompt is fine. A longer one helps because it tells the model what to animate, not because length unlocks anything.
Most people try once
1,025 accounts made at least one video. 75.7% of them made exactly one. Paying customers made 4 videos at the median, and the top tenth made 11 or more.
That gap is the main lesson for anyone building or choosing a tool. The first video decides whether there is a second. A first try that is refused, takes ten minutes or looks nothing like the photo is usually the last one. That is why failed and refused jobs on Viralance are refunded automatically, and why our measured wait times and real cost per clip are public.
Method and limits
- Source: Viralance production records, 29 January – 31 October 2026. 1,529 video jobs that finished or failed, from 1,025 accounts. Our own team, test, admin and banned accounts are excluded, as are restyles of uploaded videos and failures caused by our own bugs.
- One platform: this is what Viralance users did, not the whole market. Our users skew towards short social video, and free credits only work on the trial models.
- Small paying sample: 34 paying customers made 208 of the videos. One heavy user can move those percentages by several points.
- Shapes: a shape is recorded for text prompts. Photo jobs follow the photo, so they are not in the shape table.
- Topics: keyword matching on the prompt, multi-label, English keywords only. Prompts in other languages are counted in the totals but mostly miss the topic table.
- Privacy: we report totals only. No prompt, photo or video is quoted or shown.
More aggregate numbers from the same records are on our research page. If you want to try the photo-first workflow most people use, open the editor and add an image.
