Gemini Omni Flash: swapping the background and the actor of a real shot just by describing it

I took a video of mine, changed the whole environment and the character, and there was no roto, no tracking, no green screen: I just described the change. Gemini Omni Flash — Google’s video model, integrated into Higgsfield — keeps your footage as the base (movement, camera, action) and transforms everything else on top. And for the references, Nano Banana 2 Lite generates images at one credit, under 50 Mexican centavos. It’s the fast, cheap combination in the catalog; here’s what it’s good for and what it isn’t.

Joaquín’s pizza shot, with a neon sign that never existed on set.

The facts

• Omni Flash is a reasoning model that generates video from your input video and a plain-language prompt: change the background, tighten the shot, change wardrobe or style. Google sells it as “reshape reality.”
• Nano Banana 2 Lite is Google’s fastest, cheapest image model: one credit per image, four versions per batch, in seconds.
• It supports start frame and end frame: useful for transitions between your camera clips.
• It’s noticeably faster than Seedance — my 720 tests came out without waiting in a queue.
• The official demos are real-estate graphics, motion graphics and explainers: it understands on-screen text well.

The tests

I asked Nano Banana for a neon sign reading “Jose Pizza” on a black background. Four versions, I picked one, took it to video with the shot of Joaquín eating pizza and said: add the sign behind the character. Done. Earlier, on the same shot, I put a cowboy hat on him and turned him into a muppet — the character and the whole video, in a couple of minutes. It’s the kind of change that today costs you an afternoon in After Effects or an argument with the production designer.

The original shot turned into a muppet with a single instruction.

What I use it for and what I don’t

I’d use Omni Flash for content creation, quick things, urgent things and heavy trial and error. For something more powerful and formal — a shot that goes to a master — I go to Seedance. There’s a middle path that pays off: start in Nano Banana with the image and from there move to Omni Flash with the video.

I think there are three good video models today — Seedance, this Omni and whatever comes out of the competition between them — and in image, Nano Banana and GPT Image 2. Image, then video, on top of your own original footage, is the key to reaching a specific result. It’s not that the app does everything for you: they’re tools you use creatively, switching model and function depending on the shot.

Where I’m using it most is with the plugin inside DaVinci Resolve: there I upload my original video in 4K 10-bit and it hands back the same, with my color correction intact. For my advertising work with clients, being able to offer options and respond fast wins me clients.

Four versions of the sign in seconds: the reference that then goes into the video.

The LATAM read

It happened to me twice in the past year: we wanted to shoot on a terrace with a certain view, and from there the location was the expensive part. Or the weather we wanted didn’t show up on the shoot day and the client gets frustrated. We light, we bring the sun at will; the background is the hard thing to control. If I put an artificial sun on my character and add the sun to the background with Omni Flash, the location problem is over. For the Latin American filmmaker producing with no limits in their head but every limit in production, this is the narrative-extension tool: focus on directing and the script.

My personal opinion — Daniel Blanco A.M.C.

I tried it for the first time with you in the video and found it easier and faster than on Google’s own page. It’s cheap, it takes no time and it’s realistic in movement, which is what Google does well. It’s not the model I use for a master, but it’s the one I use to decide in ten minutes whether an idea is worth it before spending credits on Seedance. I now represent Higgsfield in Latin America and I’m taking it to universities and film schools as a VFX and narrative-extension tool — and this model is the one that convinces fastest whoever hasn’t touched it.

Try it on one of your own shots

Upload a shot you already have, ask for a background change and nothing else. If you want me to test something specific, leave it in the video’s comments: I’ll spend my own credits on your tests so you don’t spend yours on trial and error, and once we find the path, you invest.

Images: frames from Daniel Blanco A.M.C.’s video.

Original source: I changed the set and the actor of my real shot, on YouTube · Modify any video just by describing the change

Extra for you: if you’re experimenting with AI for video, my Higgsfield link has discounts and promos to get started — and using it supports this content.