4 tricks I use to get better videos from Gemini Omni every time

A few weeks ago, Google unveiled Gemini Omni at Google I/O 2026, and if you've had a chance to read my hands-on from last week, you'll already know that it left quite an impression. The model takes anything you throw at it, be it an image, text, an existing video or even an audio clip and generates a video output from it, as desired.

Now, naturally, for a model that allows for a diverse mode of inputs, the quality of what comes out the other end is heavily dependent on what you feed the model, and even more so on how you feed it to it. The single biggest shift in my experience came from realizing that Omni doesn't respond well to the same prompting instincts you'd bring to any other language or image model. To get the best out of it, you'd have to think like a cinematographer. Here's what I mean.

Give it a visual reference whenever possible

The model infers better than it imagines

Gemini Omni is a visual model, and it doesn't seem surprising that the single, most prominent improvement I saw came from giving it something visual to work with. For this test, I uploaded a simple drawing of a cartoon dog I had sketched earlier and asked the model to "bring it to life." The result was quite charming. Omni was able to transform the sketch into a fully animated character that seemed like it was pulled out of a Disney movie. It preserved all the little details, such as the floppy ears, black spots, the red collar (that I'd drawn around its neck reaching its hands), and the "personality" that such a character would have.

Of course, the difference between generating the same clip with an image versus text alone would be night and day. Asking the model to create "a bipedal dog" and "bring it to life" leaves far more room for interpretation, some of which could be horrifying and may put you off from using the platform altogether. Jokes aside, providing a specific character or visual style, especially when an idea of it already exists in your head, gives Omni something concrete to anchor itself to. Text-only prompts still have their place, but I would reserve them for the most abstract ideas.

I tried Gemini Omni, and it's so good it feels straight out of science fiction

There's a new king of video generation in town, and it's no joke

Posts 3

By Abhinav Raj

Ask to focus on subjects that matter

Less is more, especially when it comes to prompting creative models

A mistake that I kept repeating in my several attempts at experimentation is assuming that more details would produce a better output. In practice, however, it seems that Omni behaves more like a camera than a language model. Providing it with six different subjects only means dividing its attention between six different ideas.

In essence, there's certainly a sweet spot between the number of parameters you specify in your input versus the overall creative room you provide the model to generate an output. This means that, for focused work, such as a product demonstration, it helps to specify that the product in question is the primary focus of the clip, including the aspects of the product that an advertiser would want to highlight. This allows the model to understand that other elements in the background of the feed can be toned down, otherwise, they tend to serve as distractions.

Treat creative prompts like a movie shot

Direct it like a videographer, not like a search engine

A still from a video generated with Gemini Omni.

One of the key lessons I learned fairly quickly with Gemini Omni is that it's far easier (and less time-consuming) to answer important creative questions before the generation cycle begins than it is to fix them afterward. While it's tempting to generate a clip and refine it through multiple iterations, the quality of the output and your project improves dramatically when you have a clear vision of it before you reach for the keyboard.

Before writing a prompt, try to ask yourself the same questions that a cinematographer might. Where is the camera positioned? What is the subject doing? What mood should the scene convey? Is the shot wide, dramatic, or observational? The more clearly and concisely you can answer these questions, the less guesswork Omni has to do in terms of camera angles, movement, lighting and composition.

Tell it what not to do

Describing what you don't want is just as important

A subtle aspect of Gemini Omni that took me a couple of weeks to appreciate is the fact that it carries context from earlier generations within the same conversation. While this is to help maintain consistency when you're working on an active project or refining a concept, it also means that the previous outputs can influence the clips that follow.

I first noticed this when experimenting with visual styles that were not like one another. The camera angles, lighting setup and artistic choice all seemed strangely "uniform", even when I was actively trying to move in another direction. This is certainly helpful when you're building upon an idea, but it can become a hurdle when you're trying to move beyond the derivative work. As such, it becomes quite useful to introduce a set of constraints or negative prompts to retain control of the direction.

Thinking less like a prompt engineer solved most problems

Several weeks of experimentation with Gemini Omni could be condensed into a singular observation, and it's that it is always better to prompt it like a camera (or a person with a camera) rather than an LLM. Other aspects do frequently come into play, and at such times, providing the model with supporting information such as visual references, shot composition, and even negative prompting gives it a stronger creative foundation to build upon. The less time you spend asking it to make educated guesses about your intent, the less time you'll spend correcting its interpretation.