ARTEMIS LAB / BLOG
Gemini Omni: edit a video through conversation
Google announces Omni 1.2 Flash. How to plan a video test and check whether the result really shows what you intended.
Changing the background of a short video by asking for it in a conversation: that is the workflow Google highlights with Gemini Omni 1.2 Flash, announced on November 5. The model supports clips up to fifteen seconds long and conversational edits. Google’s announcement.
For a small team, the possibility worth examining is straightforward: developing an idea through small corrections. But an attractive video is only part of the job when you are showing a product. Before publishing, you need to check what the images actually communicate.
Write a brief for a tiny shoot
Imagine preparing a video to introduce a mug. You could start with one action: the mug sits on a table, a hand picks it up, and the camera moves closer to its handle. This is an example of planning a test, rather than a result we have tested.
Decide which details must stay fixed. The mug’s colour, shape and lettering matter more than a dramatic camera move. A clear brief doubles as a checklist. Without one, it is easy to judge the mood and overlook whether the product is still accurate.
Before trying the tool, write down what viewers should see, what they should understand and which details must remain accurate. Think of an invitation: beautiful decoration cannot make up for the wrong address. In a product video, attractive lighting cannot make up for an invented handle.
Change one detail, then compare
Google claims improved adherence to visual references. Access varies by product: the API is announced as generally available, while rollout in the Gemini app begins with Google AI Ultra subscribers. These are the supplier’s statements; we present no independent evaluation here. Announced features and availability.
Our suggestion for a first test is to request one change at a time. For example, change the background while keeping the action the same. Then compare the versions to see whether something essential changed along with the detail you wanted to adjust.
Keep the first version and the instruction that produced it. If each attempt replaces the previous one, working out which request helped becomes difficult. Comparing two clips side by side lets you choose for a specific reason, instead of automatically accepting the latest version because it is new.
Check the message before publishing
First watch the clip at normal speed, as someone discovering your product would. Then pause at the important moments. Check lettering, object shapes and the continuity of the action: these are suggested checks, rather than defects we have observed in this model.
Ask someone to describe what they understood without showing them your instructions. If they imagine a feature your product does not have, the video needs another edit even if it looks polished. An accurate message remains the useful standard. A convincing image can still communicate the wrong thing.
To decide whether the tool belongs in your workflow, count the attempts and manual corrections needed to produce an acceptable clip. Start with a small test and written criteria, so you can accept or reject the result without letting the visual effect make the decision for you.