"How do I make a video with Gemini?" is the question I get most, and almost everyone asks it slightly wrong. They picture typing a sentence into the Gemini chat box and getting a movie back. That's not quite how the stack works — and once you understand the actual plumbing, you stop fighting it and start getting good clips on the first try.
So let me give you the honest version, then the practical one.
First: Gemini is the brain, Veo is the camera
Gemini is Google's multimodal model family — it reasons, writes, sees images, and understands prompts. But it does not itself paint video frames. The model that turns text into moving pictures is Veo (currently Veo 3.1), Google's dedicated video generation model. When you "make a video with Gemini," what's really happening is: Gemini interprets your intent, and Veo renders the actual footage with synchronized audio.
This matters because it tells you where the quality comes from. A great clip is 70% prompt and 30% model. Veo 3.1 is genuinely excellent — native 1080p, 8-second clips, real synchronized audio (not a muted loop you score later). Your job is to feed it a prompt it can execute.
The three ways to reach Veo
There are exactly three paths, and they trade off price against friction:
- Gemini app / Google AI Studio. Free tier exists but is rate-limited and watermarked, and Veo access comes and goes depending on your plan and region. Fine for a one-off experiment.
- Google Cloud Vertex AI. The raw API. Full control, no watermark — but you need a GCP account, billing set up, API keys, and you're metered per second with the usual cloud-bill anxiety. Overkill unless you're building infrastructure.
- A hosted tool that wraps Veo server-side. You type a prompt, it generates, you pay only for what you render. No keys, no GCP, no surprise invoice. This is the path I built GeminiOmni's text-to-video tool around, precisely because options 1 and 2 are either too limited or too much ceremony for someone who just wants a clip.
Pick based on how often you'll do this. Once a month → the Gemini app is fine. Once a week or more → a hosted wrapper saves you both money and the GCP headache.
The actual steps (text-to-video)
Whichever path you choose, the mechanics are the same:
1. Write your scene as a prompt. One or two sentences describing subject, action, setting, and mood. More on structure below.
2. Pick your aspect ratio. 16:9 for YouTube and landscape, 9:16 for Reels/TikTok/Shorts, 1:1 for feed posts. Decide before you generate — re-cropping a generated video loses quality.
3. Generate and wait. Veo takes roughly 30–90 seconds for an 8-second clip. This is rendering, not a chat reply; it's normal for it to feel slow.
4. Review the audio. Veo 3.1 generates sound with the video — footsteps, ambient room tone, the right kind of music. If the audio is wrong, that's usually a prompt problem, not a model failure (see below).
5. Download or iterate. If 80% is right, tweak one variable and regenerate rather than rewriting the whole prompt.
The prompt structure that actually works
After a few hundred generations, here's the skeleton I use every time:
[Shot type] of [subject] [doing action], [setting], [time of day / lighting], [mood], [audio cue].
A concrete example:
Cinematic crane shot of a Berlin S-Bahn pulling into Hackescher Markt station, golden hour, warm and nostalgic, ambient jazz and the hum of the platform.
Notice what's in there:
- A camera instruction ("cinematic crane shot") — Veo respects shot language. "Close-up," "wide establishing shot," "tracking shot," "drone shot" all change the result dramatically.
- A subject doing a specific action — verbs matter more than adjectives. "Pulling into" beats "a train at."
- Lighting and time of day — "golden hour," "neon-lit night," "overcast morning." This single phrase controls 50% of the mood.
- An explicit audio cue — because Veo generates sound, telling it what to generate ("ambient jazz," "rain on a tin roof," "no music, just wind") is the difference between a clip that feels finished and one you have to re-score.
Things that hurt your prompt: stacking ten adjectives, contradictory instructions ("fast slow-motion"), and abstract concepts with no visual anchor ("a video about hope"). Veo renders what it can see. Give it something to see.
Common mistakes, and the fixes
"My video ignored half my prompt." You probably over-described. Veo has an 8-second budget; it can't show a full story arc. One action, one moment. Cut the second half.
"The motion looks fake / floaty." Add a physical anchor — ground contact, weight, a real-world reference ("a heavy oak door swinging shut"). Abstract floating subjects are where every video model still struggles.
"The audio is wrong or absent." You didn't specify it. Add an explicit audio clause. If you want silence, say "no music, ambient only."
"It's not photorealistic enough." Add "shot on 35mm film," "shallow depth of field," or "4K cinematic" — and lower your expectations on faces in motion, which remain the hardest thing for any 2026 video model.
Image-to-video: the underused shortcut
If you already have an image you love — a product shot, a piece of art, a character — don't describe it from scratch. Feed the image directly and prompt only the motion. This is what people searching "画像を動画に" or "make my photo move" actually want, and it's almost always higher quality than pure text-to-video because the model isn't guessing the composition. You hand it the frame; it only has to animate it. Most hosted tools, including ours, expose this as a separate image-to-video mode.
So — what's the fastest path to your first clip?
If you want to experiment once and don't mind a watermark, open the Gemini app and try Veo there. If you're going to do this regularly — for content, marketing, client work — skip the GCP setup entirely and use a tool that wraps Veo server-side so you only pay per render. That's exactly the gap GeminiOmni fills: type a prompt, get a 1080p clip with real audio in about a minute, no keys to manage and no monthly subscription burning down whether you use it or not.
The model is already good enough. The thing standing between you and a clip you're proud of isn't Gemini or Veo — it's the prompt. Start with the skeleton above, change one variable at a time, and you'll be surprised how fast you get fluent.
