I tested a brilliant AI video generator
Marco van Hurne
I build Agentic AI companies | Data Science Strategist @ Beyond the Cloud | Data Strategy Certified | AI Compliance Officer Certified
When OpenAI announced its new Sora AI video generator, and I saw the first glimpses of what it could do, I got incredibly excited and extremely disappointed at the same time. Sora is clearly an amazing piece of technology, but it's not available to the general public. And there's no indication of when OpenAI will release it. Already, OpenAI has disclosed that many scenes in Airhead had to be extensively edited using traditional tools in order to be usable.
Now, a tiny San Francisco-based AI video startup just beat OpenAI to the punch.
Last week, Luma AI released Dream Machine. It is a free AI video generator that anyone can access and use. I spent the last several days testing it, and the results are nothing short of mind-blowing.
AI video generation is here. It's powerful, it's free — and it will change social media forever.
Before we start!
If you like this topic and you want to support me:
But before I get into the nitty-gritty of Luma's Dream Machine, let's talk about how AI video generation actually works. It's a fascinating process that involves a lot of complex algorithms and machine learning.
Essentially, AI video generators like Luma and Sora use a technique called "generative adversarial networks" or GANs. These networks consist of two neural networks - a generator and a discriminator - that work together to create realistic videos.
The generator network creates fake videos, while the discriminator network tries to distinguish between the fake videos and real ones. Through this process of competition and feedback, the generator network learns to create increasingly realistic videos that can fool the discriminator.
It's a mind-bending concept !
But the results speak for themselves. Just take a look at the incredible AI-generated commercial that Coca-Cola recently released, based on Sora. It's a stunning example of what's possible with this technology.
But back to Luma. How does their Dream Machine work?
Using Dream Machine is incredibly simple. You go to the Luma website, create an account, and type in a textual prompt. Luma's Dream Machine then sets about turning your idea into an AI video. Just like you do with Dall-E, Leonardo or Midjourney for image creation
For free users, the process can take several hours — Luma is understandably overflowed with requests, much as OpenAI was in the early days of ChatGPT. But the company also offers paid plans that start at $29 per month, though. These allow for faster generation times, reducing the process to 5–20 minutes.
When Luma's Dream Machine finishes processing your prompt, the result is a full-resolution AI-generated video that you can play in your browser or download.
Like many current-generation video creators, the system likely works by first creating an AI-generated image, and then evolving that image by building it into a sequence. That image sequence becomes your video.
The upside of the approach is that you also have the option to type in text describing the video you want to see, or to prompt the system with your own image — AI-generated or otherwise. Luma Dream Machine will take the image, along with an additional textual prompt, and turn it into a video.
Luma's videos are 4–6 seconds long by default. Once you've created a video, you can extend it by typing in an additional prompt to add more scenes.
Putting it to the test
So, what kind of results can you expect from Luma's Dream Machine? Well, I put it to the test, and let me tell you, the results blew my mind.
I started by prompting Luma with existing images. I used an AI image of my Dachshund "Sloebie" and then asked Luma to turn it into a panning shot.
Prompt:
"A dachshund with a superman cape is jumping through a hoop over and over".
The result is quite realistic. Luma nails both the visuals themselves and the shaky, handheld aesthetic of a social media video. But to be honest - this would not pass the deep-fake test.
I then turned to that other great bastion of social media content: famous people.
Instead of prompting the system with an existing image, I simply asked Luma to create a close up of the face of Albert Einstein.
Prompt:
"Create a close up of the face of Albert Einstein. He is smiling. His relativity formulas are buzing above his head".
See the video
Ok-ish!
As I write this, the video is already racking up Likes and plays from real people.
Those examples were cool, but I decided to try something more creative. I asked Luma to create a video of a man in the 1910s using a laptop.
Prompt:
"Create a video of a man in the 1910's, playing around with a laptop. Slapstick"
It is a nice little clip, with a cinematic, slapstick like character. I think it nailed it with this one.
领英推荐
Overall, I was pretty satisfied with the results of my testing with Luma's Dream Machine. Particularly for a freely available tool, the videos are remarkably realistic and true to the prompts I gave it.
In particular, Luma's Dream Machine feels like a major evolution from previous video generators, like a popular tool from Runway ML. Runway can also be prompted to create videos, but they're nowhere near the quality that I saw from Dream Machine.
I can't evaluate Dream Machine against the apparent gold standard of OpenAI's Sora because, for all intents and purposes, the latter tool does not yet exist.
Yes, we've all seen the incredible demo videos from OpenAI, including their deeply narrative and beautifully produced Airhead short film. But it's unclear to what extent OpenAI is showing Sora's typical output, or how much they're cherry-picking the absolute best results to show to the public.
It's also unclear how many attempts it took for filmmakers to get usable results out of Sora.
Already, OpenAI has disclosed that many scenes in Airhead had to be extensively edited using traditional tools in order to be usable.
It's possible that the tool isn't nearly as powerful as we've been led to believe. In any event, no one can verify Sora's results, because almost no one has been allowed access to it.
And that brings us to the elephant in the room:
Why hasn't OpenAI released Sora to the public yet?
Well, there are a few possible reasons.
First, it's possible that Sora simply isn't ready for prime time yet. As impressive as the demo videos are, it's possible that the tool still has a lot of kinks to work out before it can be released to the masses.
Second, OpenAI may be concerned about the potential for misuse. As we've seen with other AI tools like ChatGPT, there's always a risk that people will use these technologies for nefarious purposes, like creating deepfakes or spreading misinformation.
Third, it's possible that OpenAI simply wants to keep Sora under wraps for now to maintain a competitive advantage. After all, if they release the tool to the public, other companies could quickly reverse-engineer it and create their own versions.
But regardless of the reason, the fact remains that Sora is not yet available to the public. And that's where Luma's Dream Machine comes in. Dream Machine certainly isn't perfect, but it's here's here now. After using it for a few days, I can already see its power as a tool for video creation. If your goal is to create short video clips destined for social media, Luma is ready for production use today.
That effectively means that AI video generation has arrived.
Shortform Internet video is today's dominant format — people watch hundreds of millions of hours of short-form videos on platforms like TikTok, YouTube Shorts, and Facebook every day.
And I will use it to post content on LinkedIn. See if that gets heads turned.
By conquering this format, Luma has effectively made AI-generated video into a reality, Sora or not.
That's exciting, but also scary.
Deepfakes
AI-generated images already pose major risks, including the easy creation of misinformation and deepfakes. But AI-generated video has a much greater capacity to distort reality.
Photoshop has been around for decades now. The majority of people know that images they see online may be altered or faked. The technology to quickly and easily conjure up a fake photo has existed since well before the AI era.
Videos are different, though. Previously, faking a detailed and visually accurate scene would have required hours of time and thousands of dollars of production costs. Internet security company Kaspersky estimates that one minute of deepfake video created using current technologies costs up to $20,000 to produce.
Until now, fake video was thus the purview of governments and institutions embarking on misinformation campaigns, not people sitting around in their living rooms.
With tools like Luma Dream Machine, that's not the case. In a few minutes and for almost no cost, anyone can now create convincing AI videos of nearly anything.
Luma does have safeguards, but the cat is out of the bag — Luma's power shows that AI video generators are here, and open-source Dream Machine clones will undoubtedly emerge soon, ready and willing to churn out deepfakes and nefarious content.
That's a big risk. But I see an even bigger one.
The biggest risk from tools like Dream Machine is not that they'll sway elections or ruin democracy, but that they'll slowly poison our social media accounts with mundane, convincing fake video content until nothing we see online is real anymore.
Funny dog videos, delicious-looking food, and even historical clips — these can now be easily and cheerfully faked using Dream Machine and its ilk.
Now that Dream Machine is freely available, you can expect to begin seeing huge amounts of convincing, AI-generated videos flooding social media. Some scrupulous creators will disclose their use of AI, but most won't.
Given how convincing these fake videos look, millions of people will be fooled by them. As the tech advances, it will be increasingly impossible to distinguish a generated video from a real video, much as it is now with AI-generated images.
And as the tools get cheaper and faster, it will be easier to fake a video with AI than to shoot a real one with your phone — much less a professional camera.
For anyone consuming online content, the message is simple: now that tools like Dream Machine are here, you can no longer trust that any video you see on social media or any online platform is real.
So, what can we do about it? Well, as with most AI content, the only response is to increase our own media literacy. Knowing that tools like Dream Machine exist, we all need to approach anything we see online with a much heavier dose of skepticism.
And we need to teach the most vulnerable among us — children, for example, and people with limited tech knowledge — how to take a similarly critical approach to video content they see online.
As the famous saying goes, "A picture is worth a thousand words, but a deepfake is worth a thousand lies" - We need to be vigilant in the face of this new technology.
With tools like Dream Machine, maybe we can all be AI-generated video stars, if only for a few seconds at a time.
What do you think? Are you excited about the possibilities of AI-generated video, or are you worried about the risks? Let me know in the comments below!
Happy watching, folks! And may your AI-generated videos be ever in your favor.
Well, that's a wrap for today. Tomorrow, I'll have a fresh episode of TechTonic Shifts for you. If you enjoy my writing and want to support my work, feel free to buy me a coffee ??
Think a friend would enjoy this too? Share the newsletter and let them join the conversation. LinkedIn appreciates your likes by making my articles available to more readers.
Signing off - Marco