Google's Gemini Omni Makes AI Video Way Too Easy

0:00How much skill does it take to make an AI video these days? I spent a few weeks on that question, prompting the best AI models out there. But the easiest one I came across is the one Google just released, Gemini [music] Omniflash. So in this video, I'll take you right through it, from making your first video off a single sentence to edits that look impossible. So by the end, you can make your own. The easiest way to make one is to just describe what you want in plain words and let it build the whole video for you. [music] So that's where we'll start. Everything today runs on

0:24Higgsfield, which is basically one site with Gemini Omniflash and every other big model in the same place. And the link is in the description if you want to follow along. So I opened Gemini Omniflash, and your description here is doing two jobs at once. It's building the picture and the sound together. So anything you leave out, the model just fills in on its own. That's why I never

0:41describe only the thing I want to see. For the first one, I ask for a golden retriever puppy running across a sandy beach at [music] sunset. And right alongside that, I tell it what the whole thing should sound like, the waves rolling in, a little [music] bark, a soft breeze coming off the sea. So I dropped that whole description in as one

0:56prompt.

1:04And there it is, a proper little video with the sound already sitting inside it, the waves and the bark and all. For a second one, I go with a street food stall at a busy [music] night market. And I handle the sound the same way, calling out the sizzling wok, the crowd behind it, a bit of music somewhere off

1:17in the distance. [laughter]

1:26Out comes the whole stall, flames and steam with the market playing away underneath it. A packed messy scene like this is the hardest test for the sound. These come out at 720p, which looks good enough for a quick first video, and I'll show you how to get these properly sharp later on. So that's a real AI video with sound made off a single typed [music] sentence. Making things from scratch is really just the start, though, because the same model will take a video you already have and change what [music] happens inside it just by asking. This next part is the one I use the most, and

1:51it only works if you write the instruction the right way. What makes it work is you start from a real video you filmed yourself and just add to it instead of building the whole thing from scratch and hoping it comes out looking like yours. You write it in two parts. First, you tell the model everything that has to stay exactly the same, the lighting, the camera, the timing, all of it. And then right at the end, you tell

2:08it the one thing you want changed. Locking all that down first is what stops it redrawing the whole shot, so it only changes [music] the thing you asked for and leaves everything else alone. I take a plain video of my hand opening from a fist into a flat palm. I lock the whole shot down and then I ask for one thing, a small glowing jellyfish that appears in my palm right as my hand

2:25opens.

2:33The hand in the room are left completely alone and the jellyfish is just there hovering over my palm appearing exactly as my fingers open out. That timing works because I told it when to happen and tied the change to a movement that was already in the video and that kind of control is normally an editing job. Next, I go smaller, a hand tilting a coffee mug towards the camera and the only thing I'm touching [music] is what's inside the mug. I lock the shot and paste in a version that turns the

2:55coffee into a galaxy.

3:05The mug in my hand never moves, but the coffee is now a slow spinning galaxy with little points of light floating up in the steam. What matters most in that prompt is telling it the galaxy is there from the very first frame and not to fade it in because by default the model eases these effects in and that soft

3:18fade is what makes an edit look cheap. Then to push it further, I try a rotating pen and ask it to keep changing as it turns from a plain pen to a candy cane stripe to a gold wand to a wooden [music] paintbrush. Same two-part setup, I just feed it the four looks in order.

3:38It walks through all four looks in time with the spin, the hand and the speed never changing, just the object in my fingers. Four changes back to back is usually where these break down and the only reason it holds is that each new look is tied to the turn of the hand. For the last one in this batch, a hand reaches in and touches a small potted plant, and the second it makes contact, the plant explodes into fast growth. And right on the touch, vines [music] and leaves come spilling out across the desk, and flowers open up with a soft

4:01rustle laid in underneath.

4:10Pin the effect to [music] the exact frame my hand lands on the leaf, and your eye reads it as cause and effect. So, now you can reach into footage you already have and rewrite what's inside it with a sentence. And everything you didn't mention stays exactly as you shot it. But, so far the camera's been pointed at objects on a desk, and the next set of edits puts me right in the frame. Doing this to myself brings in a problem the objects never had. I have to keep looking like me while everything around me turns completely mad. So, before I record anything, I give the model a reference sheet of me plus one

4:36still of the room, so it's always got something to match me against. The face shot just needs to be clear and straight on and well-lit. And the better that one photo is, the less my face changes later. So, it's worth nailing before you try anything clever. First up, I film myself sitting at my desk, just glancing around the room. Lock it and ask for one change. The whole place filling up with

4:53raccoons. I'll paste this one in.

5:00They're everywhere. On the desk, along the shelves, one sitting on my shoulder, all lit by the same lamp, so they sit in the room properly, and I haven't moved a muscle. The raccoons only look right because the prompt keeps them on real surfaces, so none of them come crawling out of the framed photos on the wall. Next, I stand up for a full-body shot and ask my outfit to snap through four looks in time with a small movement of my weight. From my hoodie to a sharp suit, to a suit of knight's armor, to a full astronaut spacesuit. [music]

5:23I run the outfit prompt.

5:32[music] It snaps straight from one outfit to the next with no fade between them, and I keep my own face and glasses right through the hoodie and the suit until the helmet covers me for the armor and the spacesuit. Then, I stand off to one side, look at the empty space next to me, and ask for a glowing portal to open there with a single Roman soldier

5:46stepping out looking completely lost. Always add the line that there is only one of you in the scene because the moment you ask for a portal and a new character, the model will sometimes clone you by mistake.

6:03The portal opens right where I looked. The soldier steps through with a slice of ancient Rome behind him. And the only thing that slipped is my chair vanishing halfway through, which one more run would clear up. For the last one, and this is the one I chose someone if I wanted to properly freak them out. I stay sitting while a second copy of me comes crashing in through the window

6:19behind me. I run the last one.

6:29A second me lands, brushes the glass [music] off, and throws a thumbs up matching my face and my hoodie spot on. Though the glass doesn't hit the jumper right and is brushed off unrealistically. So, that's one line I'd tweak and run again. And every one of those is still me, my own face and glasses filmed in the room I use every week just doing things that never happened. [music] Now, everything so far has come out at 720p, which is fine while you're testing. But the second you want to post one of these properly, you'll want it sharper. [music] And that's the one thing this model won't

6:53do. That's where a second model on the same site comes in. That second model is SeaArt's 2.0, and it runs the very same edits at higher quality right up to 4K if you want it. [music] I take four of the edits from before and run them again, word for word the same prompts, changing nothing but the model, swapping flash for SeaArt's 2.0 at 1080p. One thing to know before you do this, SeaArt treats your original video more as a loose reference [music] than something to edit frame by frame. On SeaArt, the outfit switch came out best of the lot, smoother changes and tighter timing than the flash version. Close behind was the

7:21coffee galaxy, already strong on flash and a bit sharper here. Where it struggled was the plant bloom and the Roman portal. Both worse here and the portal basically unusable, though a couple of prompt tweaks would probably sort it. So, SeaArt gives you the higher quality right up to 4K while flash gets you a strong edit for a fraction of the cost. For something going on a big screen or a paid job, that extra quality is worth it. But for anything headed to social, it barely shows once the platform compresses it. So, most of the time flash is plenty and that cost is a bigger deal than it sounds. Flash is

7:49built to be quick and cheap and edit runs you around 24 credits, where a standard C-Dance generation is about 96. So, for the day-to-day stuff, flash is the obvious pick. So, the rule I stick to is dead simple. Build and test everything on flash while it's cheap. Then only rerun the two or three shots you're keeping on C-Dance for the quality and you end up with sharp final videos without wasting a load of credits. And both models run on the same Higgs Field account and use the same credit balance. So, switching between them is just changing the model in a drop-down. And that second copy of me

8:14crashing through the window came from nothing more than a written description, same as every other edit here. And those raccoons on my desk and the galaxy in my mug both started as a single line you can copy today. You've already got the exact prompts. The only thing missing is the one place that runs every model off a single balance. [music] Open Higgs Field from the first link in the description and make your first one right now. Thanks for watching and I'll

8:34see you in the next one.