Google tools · Flow
Google Flow: a real place only from a real photo.
Google Flow makes 8-second clips with the Veo 3.1 model. From text alone it made an excellent clip of a city that does not exist. From a real photo it showed the real place. In two evenings we made 9 clips and spent 548 credits. Two clips run on a website with an "AI video" label. We made the third video without Flow, from real drone footage, for 0 credits.
The numbers
| What | Value | Where the number comes from |
|---|---|---|
| Clips | 9 × 8 s | 1080p, with sound; 5–6 Nov 2026 |
| Still images (Nano Banana in Flow) | 4 | not charged |
| Credits in total | 548 | 4 × 12 + 5 × 100 |
| Daily credits on the account | 50 | Flow's message at the first generation |
| Price per clip | 12 credits · 100 in Quality mode | the approval prompts in Flow |
| Generation time | ≈ 1–2 min per clip | from the log timestamps, not measured |
| Text-only clip | rejected | the place was invented |
| Clips from a real photo on a website | 2 | labelled "AI video" |
| Video from real drone footage | 0 credits | edited on our own machine, 14.9 s on the site |
Google's plans and prices change; these are only the credits Flow showed us on 5–6 Nov 2026.
The lesson that cost us one clip
The first clip was text only: a building in Varna at sunset. The quality was excellent, but the buildings, streets and coast were invented. For a real place that is an error, not a style.
The fix is simple: a real photo as the first frame, and the prompt describes only camera movement, light and sound. Plus, explicitly: "do not invent or change any buildings".
The four modes we used
- Text → video. Fine for abstract scenes. Not for a real place.
- Frame → video. A real photo as the first frame. This is how the clips on the website were made.
- First and last frame. Two drone photos: start and end. Flow makes the smooth flight between them. It does not accept three photos as exact frames.
- Image → video. First a still image with the exact number of objects, a human check, then motion.
What Flow cannot do (and how we work around it)
- It does not count. We asked for three fish and got 4–5. Even from an image with exactly three, Veo added a fish after 2–5 seconds. Fix: image with the exact count → video → frame-by-frame check → keep the clean part.
- It does not keep the camera still. "Camera locked", and it still zooms in towards the end.
- A logo is not described in words. We described our mark in words and got the wrong shapes. A logo is given as a file; we made the right clip locally from the file, for 0 credits.
- Sometimes it says "generating" and generates nothing. We check that a new clip really appears in the project.
- Credits run out fast. The daily 50 are enough for 4 clips at 12; one Quality clip is 100. We turn off automatic approval and approve every paid generation separately.
- Sound is not visible in frames. We check it in the browser with sound on.
How we show it on a website
The clip starts on its own, muted, plays once and stops on the last frame. A photo stands in while it loads. Sound is switched on with a button. Every clip with generated frames carries an "AI video" label in the corner; an edit of real footage has no label. The file went from 14 MB to 3.2 MB.
If real footage exists, we edit it first. AI is for what is missing.
The recipe
The prompt, the exact steps and the checks are on Better Mechanics, our recipe library:
Open the recipe →Frequently asked questions
Can Google Flow make a video of a real place? Yes, if it starts from a real photo as the first frame. From text alone Flow invents the buildings and streets.
How many credits does a Flow clip cost? For us 12 credits per clip and 100 per clip in "Veo 3.1 - Quality" mode, with 50 daily credits. In two evenings we spent 548 credits on 9 clips.
Can Veo show an exact number of objects? Not reliably. We asked for three and got 4–5. We make an image with the exact count first, then the video, and cut the clean part.
Should AI video be labelled? Yes. We put "AI video" on every clip with generated frames. An edit of real footage has no label.