What would you create if you had tools that offered endless possibilities at your fingertips? Perhaps you would create a masterpiece. Or you might freeze, overwhelmed by being put on the spot by technology, before coming up with the worst video ideas ever.
Perhaps that is just me though. Google has just unveiled Veo 3, its latest and most advanced video-creation model, for users in Ireland. It is available with a Google One AI subscription, which is €22 a month for Google AI Pro, which also includes Gemini in Google’s productivity and email products, 2TB of storage in Google Drive, photos and email, and 1,000 AI credits to use as you see fit, including creating videos.
The thought of it was very exciting: a tool that would spit out realistic video clips from a few text prompts. The reality was that I wasted my first few credits on nonsense.
One of the biggest concerns about the growth of AI was that it could possibly take over the jobs of humans. The second biggest concern was that AI-generated video and deepfakes would become more common.
Will the newest round of AI tools help assuage those fears? For the most part, no.
Because Veo 3 isn’t just about creating simple clips. Building on the capabilities of Veo 2, the clips are cinematic, the model generates audio, such as ambient noise, and creates dialogue for on-screen characters. The technology understands camera angles and framing; you can tell it to use a specific look or a lens style. It is all the stuff that you would have expected to need human input for.
But the clips are limited to eight seconds in length. That will mean a lot of prompt generation for something that wouldn’t even make a reasonable social media clip.
When I finally came up with something I wanted to create, it was an easy process. Veo takes one to two minutes to generate each clip, unless you have given it some detailed instructions.
The first was a video of a backyard children’s party at dusk, with a tepee tent, bunting and small children playing.
With some input from the younger members of the family, the second was a cartoon-style vegetable wedding – the carrots were in tuxes, the guests were fancily dressed broccoli and potatoes. The dramatic hook: an objection to the proceedings that halted the wedding. The eight seconds had the kids hooked.
Google is trying to tread a fine line here. Veo’s output is realistic enough to fool you most of the time. Occasionally though, you still need human input. Like most AI, Veo is not perfect. The beach video, for example, started with waves engulfing the couple sitting on the beach, receding to show perfectly dry people sitting a few feet away from the waterline. The IMDB goofs section for this particular movie would have a field day.
The child’s party video had a glitch that saw limbs separate weirdly; the cartoon carrots sounded fine, but the accompanying subtitles – unasked for – didn’t make sense.
You need to be specific about the soundtrack too. If you want ambient noise, say it. The first attempt at the restaurant scene was the quietest I’ve ever heard a full establishment. No one spoke, and you could hear cutlery being used in a way that would usually have blended into the background noise. Adding “realistic restaurant background noise” made a difference. This is why prompt engineering is a specialist field all of its own.
There are some things Veo simply refuses to do, for obvious reasons. Unless you have the patience to try to figure out a way around Veo and Gemini’s restrictions, you are bound by the policies of the service.
It wouldn’t create a video of a certain controversial tech mogul dancing around a campfire, timing out after 10 minutes of waiting. When it came to political figures in Ireland, Gemini refused to create a video of the Taoiseach wearing a Santa hat at a mock Christmas party.
[ New Bill targets AI ‘deepfakes’ and identity ‘hijacking’Opens in new window ]
Depending on what subscription you have, you will also find yourself limited to certain number of video creations per month.
In Flow, Google’s AI studio that is strictly speaking not yet available in Ireland, I created a fly-by of a dystopian cityscape, with decaying buildings and abandoned cars, accompanied by a soundtrack using Veo 3. It created two versions by default; you can up that to four. They all hit the mark on visuals.
But the real test was in the most realistic video I created – a black cat, slinking through the grass, stalking prey while purring. It looked remarkably like our family pet, so I sent it via WhatsApp to a few family members.
Not one asked if it was AI. Task accomplished.
Good
If you have a specific scene in mind, Veo can help bring that to fruition. The trick is knowing what you want and being very specific about it. The prompt is key. The more detail you give Veo, the more likely it is that you will get what you want.
Bad
Occasionally it generates some flawed AI slop, such as a magic wave that miraculously leaves everyone dry in its wake.
Everything else
You have to stick to the rules laid down by Gemini. Deepfakes are out, as is creating content the system flags as offensive.
Verdict:
AI-generated content is getting easier and more convincing. Veo 3 makes another big leap on this front.
