Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI.
Today, the promising AI video startup Runway ML introduced its newest foundation model, Gen-3 Alpha. This model enables users to create high-quality, ultra-realistic scenes of up to 10 seconds with diverse camera movements, using only text prompts, still images, or pre-recorded video.
We had a chance to speak with Runway’s co-founder and CTO, Anastasis Germanidis, about Gen-3 Alpha and its significance in the rapidly evolving and competitive video AI software industry. Germanidis shared insights into Gen-3 Alpha’s distinct features and how Runway plans to compete in the market.
For Gen-3 Alpha’s rollout, it will first be available to paid users in the coming days, followed by free users.
VentureBeat: Can you explain how Gen-3 Alpha differs from previous versions like Gen-1 and Gen-2? What are the key improvements?
Anastasis Germanidis: Since we released Gen-2 over a year ago, we’ve learned a lot. Initially, Gen-2 could only be prompted in a very basic way with text. Subsequently, we added extensive controls for camera and object motion, which became crucial for users. For Gen-3 Alpha, we invested significantly more time and resources, especially on data captioning during training. Now, users can prompt it with more complex interactions, including camera movements in different styles and character interactions.
Another major upgrade is the scale of computation. Increased computing power enhances the model’s ability to maintain geometric consistency of objects and characters, preventing them from morphing over time, which was a challenge with earlier models. We also noticed that prompt adherence improved significantly with the scale.
Additionally, Gen-3 Alpha is built as a modular framework, allowing for easy integration of features like image-to-video and camera control. This structure lets us rapidly introduce new tools, and users will see a host of new capabilities built atop Gen-3 in the coming months.
VentureBeat: Given the recent launch of competitive models like Kling from China and Dream Machine from Luma AI, how does Runway view competition and user excitement about these new offerings?
Germanidis: While Gen-3 has been in development for a while, staying at the forefront is crucial for us to provide outstanding results for users. The base model isn’t a long-term differentiator; it’s essential for subsequent tools and controls. What sets us apart is our extensive toolset, fine-tuned with years of user input, embedded into Gen-3. This extensive infrastructure is something new models lack.
VentureBeat: Where do you see the market for video AI headed? Will there be multiple AI tools for different purposes, or a dominant player?
Germanidis: In the long term, we expect photorealistic video generation models to become standard. The differentiator will be how effectively these models are used and the community built around them. We view a new genre around AI film as emerging. Our focus is on creating controllable storytelling tools for professional creators, leveraging our extensive community and toolset. This approach will differentiate us in a market likely to host multiple niche players rather than a single dominant entity.
VentureBeat: Can you tell us more about collaborations with artists and media companies in developing Gen-3 Alpha?
Germanidis: We have an in-house creative team working closely with our researchers. We’ve also collaborated with media and entertainment companies for custom models tailored to specific needs. While we’ve not publicly disclosed most partnerships, one example is Getty Images. Custom models offer improved performance for consistent character generation and specific stylistic requirements.
VentureBeat: How does the training process for Gen-3 Alpha differ from previous models?
Germanidis: Typically, video generation models are trained on paired text and video data. Our approach involved highly detailed and numerous captions throughout the video’s course, offering precise scene definitions and changes. This detailed annotation was done by our internal team.
Regarding training data, we use proprietary data sources, and while we have various data partnerships, we haven’t disclosed specifics. This curated data helps us maintain high-quality model training.
VentureBeat: When can users expect to access Gen-3 Alpha?
Germanidis: Gen-3 Alpha will soon be available in the Runway product. We’re currently ensuring we have the capacity to serve the model, starting with paid users and rolling out to free users shortly after.
VentureBeat: How do you see Gen-3 Alpha being used by professionals versus independent creators?
Germanidis: Usage spans a spectrum. Professional filmmakers have used Gen-1 and Gen-2, primarily for supplementary footage or storyboarding. With Gen-3 Alpha, we expect broader production use, but it will be a gradual transition. The advantage of AI tools is the speed of iteration, enabling quick idea realization without extensive real-world planning or travel.
VentureBeat: Do you foresee fully AI-generated films becoming the norm?
German