The creators behind OpenCV, the world’s leading computer vision library, have launched CraftStory, an AI video startup aiming to surpass rivals like OpenAI’s Sora and Google’s Veo. Their new Model 2.0 can generate realistic, human-centric videos up to five minutes long, a significant leap in duration for the AI video industry.
Key Takeaways:
- CraftStory, founded by OpenCV creators, launches Model 2.0 for AI video generation.
- The system produces videos up to five minutes, significantly longer than competitors.
- It utilizes a parallelized diffusion architecture, contrasting with sequential methods.
- Focus is on enterprise applications like training and product demonstrations.
CraftStory’s Breakthrough: Long-Form AI Video Generation
CraftStory’s Model 2.0 addresses a critical limitation in current AI video technology: duration. While OpenAI’s Sora 2 is capped at 25 seconds and other models at 10 seconds, CraftStory can produce continuous videos up to five minutes. This extended length is crucial for enterprise use cases such as training modules, marketing campaigns, and detailed product demonstrations, where shorter clips fall short.
Victor Erukhimov, CraftStory’s founder and CEO, explained that the system is designed to better adhere to user instructions, a common frustration with existing AI video tools. “We developed a system that can generate videos basically as long as you need them,” he stated.
The Technology: Parallel Diffusion Architecture
The innovation lies in CraftStory’s parallelized diffusion architecture. Unlike traditional sequential methods that build video frame-by-frame, CraftStory runs multiple smaller diffusion algorithms simultaneously across the entire video duration. Bidirectional constraints allow future frames to influence earlier ones, preventing the propagation of artifacts and ensuring temporal coherence.
Crucially, CraftStory trained its model on high-quality, proprietary footage shot with professional actors and high-frame-rate cameras. This focus on data quality, rather than just quantity or computational power, is cited as a key differentiator. Currently, Model 2.0 is a video-to-video system, allowing users to animate a still image using a