I made 25 videos using Google’s Veo 3. Here’s how it went.

When the AI-generated “Will Smith consuming spaghetti” video went viral a bit of over 2 years in the past, I wasn’t as skeptical as some about the way forward for AI video era. I anticipated enhancements, however I by no means imagined the know-how would advance so shortly. Certainly, it was simply final month that Google rolled out Veo 2, its second-generation AI video generator mannequin, to the general public, and the corporate is already again with a way more spectacular mannequin. After making over 25 movies with it, I’m satisfied that Google’s Veo 3 is a mind-blowing development in AI video era, for higher or worse.

What’s Veo 3?

Veo 3 is Google’s state-of-the-art text-to-video era mannequin. Like Veo 2, Veo 3 creates high-quality movies in a spread of topics and types, even capturing nuanced object interactions and human expressions. Each fashions additionally block “dangerous requests and outcomes” and mark their video outputs with an invisible watermark referred to as SynthID.

The Veo 2 mannequin may solely produce silent movies, making it extra like a high-quality GIF generator. The brand new Veo 3 mannequin, nonetheless, helps native audio era, placing it leaps and bounds forward of its predecessor. The brand new mannequin cannot solely generate sound results and ambient noise but additionally create dialogue that’s synced with the video.

Whereas Veo 3 outputs are nonetheless restricted to quick, 8-second video clips, the addition of native audio era has allowed individuals to create some truly mind-blowing AI videos which have taken the Web by storm. I’m positive you’ve seen a few of these movies already, but when not, we’ve put collectively a set of over 25 movies made by Veo 3 that show the device’s prowess and its present limitations. Whereas Veo 3 generally is a ache to work with, its low barrier to entry makes it an unimaginable device for anybody with sufficient time to create convincing, life-like movies, and I’m not satisfied the world is prepared for this.

Veo 3 nearly makes it too simple to create life like movies

If you happen to’ve spent any period of time on social media in the previous couple of weeks, then you definitely’ve in all probability seen individuals argue over whether or not 100 males may beat 1 gorilla in a struggle. It’s change into considerably of a meme, with laypeople and specialists alike chiming in on the talk. Some beginner video makers have even created their very own simulations of the hypothetical brawl. I wished to see how simple it will be for me, somebody with nearly no 3D animation expertise, to make a video displaying 100 males tackle 1 gorilla.

It was so simple as asking the Gemini chatbot, “Create a video displaying 100 males combating one silverback gorilla.”

Now, I’m positive in case you pixel-peep, you’ll find some errors. Perhaps you’ll spot some males or weapons within the background showing or disappearing randomly, or maybe you’ll discover that there clearly aren’t 100 males on this 8-second clip. However in case you had been to easily watch this video on a small smartphone display screen, you’d be hard-pressed to search out any main points at a look.

This video positively captures the chaotic, fast-paced motion that may ensue when 100 males tackle 1 gorilla. The sound that Veo 3 generated for the gorilla’s punches had weight to it, making it really feel plausible. I knew it was an AI-generated video, in fact, as a result of I used to be the one who made it. However once I confirmed my mother — who was unaware of the memes on social media — this clip, she requested me what film it was from!

One other video that demonstrates Veo 3’s expertise at simulating animal physics is that this one:

I requested Gemini to make a video of “a bull rampaging in a store promoting positive china,” and it created a video the place, once more, in case you had been to pixel-peep, you’d in all probability discover points. However at a look, every thing appears strikingly actual: the bull’s motion by way of the store, the way in which dishes scatter, and the accompanying sound of them breaking. Most retailers promoting positive china would in all probability be extra organized, however some on the market would possibly certainly appear like this.

Whereas I feel the “100 males vs. 1 gorilla” video does a good job displaying how Veo 3 handles individuals, this subsequent instance higher illustrates its potential to seize the nuances of human expressions. One of our readers asked us to create a video of “a British Parliament debate between two males utilizing the roadman accent,” and I used to be amazed at the way it turned out.

Veo 3 generated some actually life like, delicate actions within the — want I remind you, AI-generated — palms and facial muscle tissue of the person on the left as he says, “Have you learnt what? Blud.” And the way in which the person on the suitable strikes as he coughs within the different man’s face felt extremely actual.

What I feel Veo 3 does finest is create life like movies of completely unreal conditions. Certain, you in all probability received’t ever see 100 males struggle 1 gorilla, a bull rampaging by way of a positive china store, or two British Parliamentarians arguing in a roadman accent in your lifetime, however they’re all issues that might conceivably occur.

You’ll by no means see an actual video depicting a “hyper-magnified view of a bustling ant colony” the place the ants are literally “intricate clockwork robots” constructing mini skyscrapers from “glowing sugar crystals,” just for this intricate scene to be disrupted by a human finger urgent down from above, culminating in a close-up of a spherical sugar crystal.

And also you’ll by no means see “an asteroid crashing into an ocean of water balloons.”

Each of those movies show Veo 3’s spectacular talent at simulating physics interactions, even when coping with such outlandish eventualities. To my beginner eyes, these appear like they may have been meticulously created by an skilled in 3D animation, however they weren’t.

My favourite video to return from my testing is the one the place the Google climate froggy mascot pops out of a Google Nest Hub to have a picnic on the kitchen counter. I simply love how the frog pops out of the show, shortly lays out a picnic blanket, and waves on the digital camera. I additionally love how the digital camera pans out to indicate the kitchen, and the way precisely Veo 3 rendered the frog’s shadow and reflection on the kitchen counter.

The video isn’t good, although, because it surprisingly doesn’t have any audio. Additionally, the textual content on the Nest Hub says “Nesst” as an alternative of “Nest,” displaying that Veo 3 nonetheless has points with rendering textual content precisely in movies. Once I requested Veo 3 to redo the video however add sound, the consequence was far more unsettling: As an alternative of a cute frog coming out of the show, what appears like an individual in a frog costume jumps over the show and begins a picnic on the counter earlier than saying, “Good day, all people!”

This redo encapsulates among the points that Veo 3 nonetheless has with consistency and immediate adherence. The truth is, this was simply one among many examples of Veo 3 struggling to stay to my immediate, regardless of Google’s acknowledged enhancements to immediate adherence. In a single sense, I’m nearly glad that Veo 3 isn’t good; the mannequin received’t all the time offer you what you’re searching for, and that makes it more durable for individuals with malicious intentions to abuse it.

Ought to we be glad that Veo 3 isn’t good?

The movies I shared above, in addition to the many great Veo 3-generated videos you’ll have already seen on-line, might have given you the impression that Veo 3 all the time produces glorious outcomes. It could very simply achieve this, however it nonetheless struggles so much with producing legible textual content and adhering to the immediate.

For instance, once I requested it to create a video of “a lady paragliding from the highest of the Eiffel Tower,” it gave me a fairly realistic-looking video of a lady paragliding…subsequent to the Eiffel Tower.

Or when somebody requested me to make a video “the place somebody varieties a immediate for Google Veo inside Google Veo,” the mannequin churned out a video with unintelligible textual content on a laptop computer display screen.

Probably the most amusing situation I’ve encountered with Veo 3 is its incapacity to grasp what a ‘bugdroid’ is. Bugdroid is the official nickname of the Android robot mascot, however Veo 3 constantly fails to precisely painting the robotic in its generated movies, typically creating generic robots with massive eyes or bug-like antennas. It’s not just like the mannequin refuses to generate movies that includes the Android mascot because of model security issues; it might probably simply be made to take action in case you inform it to make a “inexperienced Android bot.”

Talking of brand name security, it’s good to see that Veo 3 not less than has some primary protections towards producing movies that includes well-known individuals. If you happen to ask it to, for instance, make a video that includes YouTube star Mr. Beast, it’ll refuse to take action. If you happen to attempt to work round this by describing the precise individual as an alternative of offering their title, Veo 3 will nonetheless refuse to generate the video.

With some actually intelligent prompting, Veo 3 can positively nonetheless be coaxed into creating movies that includes well-known individuals, as we’ve already seen a number of Veo 3-made recreations of Will Smith eating spaghetti online. That is positively problematic given how life like Veo 3’s movies may be. Even movies that don’t characteristic celebrities may cause a stir, as demonstrated by a faux video of a lady being denied boarding because of her eager to convey her “emotional support kangaroo” onboard.

To fight this, Google has began to place a visual watermark on all Veo 3 movies generated by way of Gemini. Nonetheless, there’s an exception: this watermark isn’t positioned in case you’re a Google AI Ultra subscriber utilizing Flow, Google’s new AI-powered filmmaking device, to generate movies. The Google AI Extremely plan prices an eye-watering $249.99 monthly, which is costly sufficient to discourage some however not all individuals with malicious intent.

With Movement, you possibly can even information Veo 3’s output with your personal photographs or AI-generated images. This permits for larger management over the output and allows producing movies that higher align together with your artistic intent, seize your required aesthetic, and match your characters’ designs. It additionally opens up new avenues of video creation by permitting Veo 3 to generate expanded scenes, reimagine movies in several types, take away undesirable objects from movies, and animate drawn characters.

Movement mitigates lots of the inherent limitations of utilizing Veo 3 in Gemini, and I can genuinely see it being helpful for beginner and even skilled filmmakers. However it additionally circumvents Google’s new seen watermarking coverage and solely provides an invisible watermark that few platforms assist. It additionally makes it simpler to place collectively longer movies, which suggests it has the next potential for abuse.

I don’t suppose there’s any turning again at this level; AI video mills are right here to remain, they usually’re simply going to maintain getting higher. Veo 3 is leaps and bounds higher than the primary text-to-video fashions, and it’s solely been a couple of years. With how a lot present and new video knowledge Google can pull from YouTube, the corporate will undoubtedly make main enhancements in its upcoming Veo 4 mannequin.

See all of the movies we made with Veo 3

If you happen to’re curious, right here’s each single AI generated video we made with Veo 3 by way of Gemini:

• Create a video displaying 100 males combating one silverback gorilla. (Link)
  • Are you able to repeat this, however make it prefer it was uploaded to Snapchat circa 2018, filmed on an iPhone? (Link)
• Create a cinematic trailer for an imaginary sci-fi film set on a distant planet with floating cities the place the protagonist is secretly the son of the villain. (Link)
• Create an animated video of backyard gnomes developing a futuristic AI supercomputer utilizing CPUs that resemble carrots, potatoes, and broccoli. Present them working in a magical underground lab with glowing circuits and enchanted instruments. (Link)
• Make a video displaying the Android bugdroid strolling down a path by itself, taking a look at a smartphone it’s holding. That bugdroid will get stunned by a few different bugdroids which might be hanging out collectively and invite the primary droid to affix them. Every bugdroid ought to be sporting a hat that claims “Android Trustworthy”. (Link)
• A person strolling on water. (Link)
• A inexperienced Android bot messing with a purple Apple and eventually consuming it. (Link)
• Make a video of three humanoid mechs combating towards a military of 20 purple coloured humanoid mechs through aeriel fight above the skies of Tokyo. The entire combatants are actively combating and never taking turns or ready round. The digital camera is slowly zooming outwards from the motion all through your entire scene. Behind a distant cloud, a silhouette of a flying lizard monster may be seen. (Link)
• A bull rampaging in a store promoting positive china. (Link)
• An influencer proclaims to the world, through a brief kind, vertical video, a 100 man vs 1 gorilla showdown whereas displaying the contenders. (Link)
• A girl paragliding from the highest of the Eiffel Tower. (Link)
• A surreal F1 race the place iconic automobiles and drivers from completely different eras compete on a monitor that morphs by way of time and area. (Link)
• A hyper-magnified view of a bustling ant colony, however as an alternative of ants, tiny, intricate clockwork robots frantically construct and dismantle miniature skyscrapers fabricated from glowing sugar crystals. All of a sudden, a colossal, slow-motion ‘finger of doom’ descends from above, casting a large shadow, accompanied by a booming, distorted ‘THUD’ adopted by the sound of crumbling glass and the squeaking, frantic whirring of the robots as every thing collapses into glowing mud. Finish on a close-up of a single, completely spherical sugar crystal touchdown silently. (Link)
• Create a video of a basketball recreation between 5 bugdroids in yellow jerseys and 5 bugdroids in white jerseys the place one of many yellow bugdroids dribbles the ball from his three level line to his free throw line then again to his three level line then turns round and shoots the ball to the basket. The ball bounces on the rim then hangs within the air then drops in. The bugdroids that shot the ball yells in delight whereas placing his palms round his throat and his different 4 teammates wearing yellow run over to him to rejoice. Bugdroids, by the way in which, are the time period for the Android OS robotic mascot. (Link)
• Create a video of a seashore crammed with individuals with everybody doing a singular seashore exercise whereas the digital camera pans to the left 180 levels, then the digital camera pans proper again to the unique spot. Within the foreground is an individual with a surfboard strolling left, and the individual is ultimately out of the body, however will reappear as soon as the digital camera pans over him once more as it’s returning to the primary body of the scene. (Link)
• The froggy mascot coming out of a Google Nest Hub and having his picnic on a kitchen counter. He then greets the digital camera because it zooms out to disclose the kitchen. (Link)
• Create a video of the Android Bot mascot holding a Google Pixel cellphone and texting his buddy, a purple Apple holding an iPhone. The attitude ought to be over the shoulder of the Android Bot, displaying the Pixel cellphone’s display screen with the Google Messages app on display screen. The message the Bot is sending ought to learn, “Thanks for getting the message!” As he hits ship, the video ought to present the purple Apple receiving the message on his iPhone, and the message ought to be proven in a inexperienced bubble within the Apple Messages app. (Link)
• Make a video impressed by a Bollywood motion film scene that transitions to a musical dance sequence after a couple of seconds. Afterwards, the motion sequence resumes from the place it was suspended. The entire actors are of Indian descent and the lead actor is wearing conventional Indian garbs. Alongside him is his sidekick who’s dressed as a name heart employee. (Link)
• Create a scene the place a black male wearing Hawaiian clothes is operating away from a T-Rex. The 2 are transferring within the route of the viewers. This scene performs for 1.5 seconds and marks the top of section A of the video. Section B of the video now begins. The digital camera zooms in on the person’s face as he appears to the left aspect of the display screen. The scene pans to the left to disclose an in depth up shot of a white male dressed solely in a seashore towel and sun shades. The brand new character lifts up his glasses slowly and reveals a bewildered expression. This part marks section B of the video and lasts for 4 seconds in complete. Section C of the video begins at this level and the digital camera now shifts again from the white male again to the black male operating from the T-Rex. The digital camera zooms out and the T-Rex lets out a roar. Within the background, a volcano may be seen erupting as soon as the roaring begins. (Link)
• A gamer who’s dwell streaming themselves enjoying a tactical, turn-based, fantasy-themed Japanese RPG. The participant is presently in a battle with a celebration of 4, one thief, one white mage, one paladin, and one black mage, combating towards a purple dragon. The participant may be seen in a small sq. overlay on the underside proper nook of the display screen. (Link)
• An asteroid crashing into an ocean of water balloons. (Link)
• Present us a video of NASA astronauts discovering life on mars from the angle of the mission management room. (Link)
• A video the place somebody varieties a immediate for Google Veo inside Google Veo. (Link)
• Giants Causeway with the basalt columns rising and sinking into the ocean in waves. (Link)
• A battle between unicellular and multicellular organisms. Video from the battlefield. (Link)
• A British parliament debate between two males utilizing the roadman accent. (Link)