Runway announces faster video AI model Gen-3 Alpha Turbo
Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More
It’s been a while since we’ve had a fancy new AI model with the word “Turbo” to explore, but Runway is making sure it doesn’t stay that way.
The New York City-based startup that just earlier this week turned heads with its new image-to-video update for its incredibly realistic Gen-3 Alpha model today announced on the social platform X (formerly Twitter) that it is debuting yet another, faster version of said model: Gen-3 Alpha Turbo, which will be “rolling out…with significantly lower pricing over the coming days.”
In its post, Runway said that the Turbo model was “7x faster than the original Gen-3 Alpha.”
Runway co-founder and CEO Cristóbal Valenzuela also posted on X a message stating that users could generate new videos with Gen-3 Alpha Turbo in “real-time,” or at least close to it, producing a 10 second video in 11 seconds.
We previously tested the new image-to-video feature on the prior version of Gen 3-Alpha and found that it was already quite fast, generating videos from stills in often less than a minute. But apparently, Runway thought this wasn’t good enough and sought to do much better.
It makes sense if the company wants to maintain its lead in offering highly realistic, Hollywood-quality generative AI video models even as insurgents and upstarts nip at its heels, including Pika Labs, Luma AI, Kling, and of course, OpenAI’s Sora, the latter of which remains only available to a small group of handpicked testers despite being initially shown off in February.
Valenzuela also told a questioner on X that Runway was working on an update to its mobile app to include support for image-to-video with Gen-3 Alpha.
More for less?
But why would Runway offer a newer, faster version of its latest model with the same quality of AI video generations at a lower price?
Aside from potentially being a simpler and less computationally heavy model for it to run on its servers (and therefore cheaper), the company may also be banking on the fact that faster generations will lead to more overall usage and thus, more overall spending on its subscription plans or a-la-carte generation “credits” model.
Right now, Runway offers a variety of monthly subscription plans that each come with a set number of credits which must be exchanged for each generation of still images or video on its platform.
Gen-3 Alpha, the most prior model, costs 10 credits for every 1 second of video generated.
Its older Gen-2 model is priced at 5 credits for every 1 second of video, while interestingly, its oldest Gen-1 model is most expensive at 14 credits per 1 second of AI video.
It would probably make sense for the company to offer Gen-3 Alpha Turbo at around 7 credits for every 1 second of video, then, or perhaps as low as 5.
Training questions persist
Last week, 404 Media obtained a spreadsheet allegedly from a former Runway employee showing the company’s plans to scrape and train its AI models off of popular YouTube channel videos, including copyrighted content from major motion picture and TV shows that was ripped and posted or clipped by other YouTube users.
While the company faced criticism from some on the web for the tactic, Runway has not yet commented on the 404 Media report and spreadsheet.
That said, it is already facing lawsuits alongside other gen AI creative generator companies from creators for allegedly violating copyrights on their still images.
Yet scraping, as I maintain, was broadly viewed as permissible ever since Google followed a similar tactic to build its search index and sell ads against it.
Still, today, one prominent critic of unauthorized gen AI scraping, former Stability AI exec turned nonprofit founder Ed Newton-Rex, who offers paid certification for ethically trained AI through his new organization Fairly Trained, called out Runway again on X and asked for the company to disclose its training data set.
Most leading generative AI companies, even those behind open source models such as Meta’s Llama 3.1, have not fully disclosed the intricacies of their training data sets and it is reasonable to me to conclude they view the training data as a proprietary and competitive secret.
But we’ll see as these lawsuits make their way through the courts if discovery forces gen AI model providers such as Runway to disclose their training data, and if they are found to be in violation of any copyrights and if so, what the remediation or damages would be.
