OpenCV founders launch CraftStory with 5-minute AI videos

Most AI video generators today are great at giving you a five-second visual shock, but try making a full product demo or a YouTube tutorial, and the illusion instantly shatters. Even OpenAI's Sora 2 caps out at around 25 seconds.

Now, the creators behind OpenCV, the world's most widely used computer vision library, are stepping in to tackle the duration wall. Their new startup, CraftStory, just came out of stealth on Tuesday with Model 2.0, a system capable of generating continuous, realistic human-centric videos up to five minutes long. What makes this even wilder is their war chest: while OpenAI raises billions, CraftStory built this with just $2 million in backing, almost entirely from Wrike founder Andrew Filev.

Breaking the sequential generation trap

Traditional video models treat time as a linear axis in large diffusion volumes. The moment a glitch appears in second two, that artifact compounds into a nightmare by second eight, which is why everyone stops short of 30 seconds.

CraftStory founder and CEO Victor Erukhimov took a radically different route with a parallelized diffusion architecture. Instead of generating frames one after another and stitching them together, the system runs multiple smaller diffusion processes simultaneously across the full five minutes. These processes talk to each other bidirectionally, meaning what happens later in the video can actually constrain and correct earlier frames to stop errors from compounding.

Their training data strategy is equally unconventional. Rather than vacuuming up blurry 30fps YouTube clips, CraftStory rented studios and shot high-frame-rate proprietary footage of professional actors. This captured crisp details in fast-moving areas like fingers, eliminating the motion blur that usually trips up AI models.

Right now, Model 2.0 works as a video-to-video tool. You provide a static image of a character and feed it a driving video of a person moving. You can supply your own movement or pick from a library of professional actor presets, with actors earning royalties when their data is used. There is also dedicated lip-syncing and gesture alignment to keep body language in tune with voice cadence.

The catch? Speed. Generating just a 30-second low-resolution clip currently takes around 15 minutes.

Why it matters

We think CraftStory is taking a brilliant commercial shortcut. By focusing strictly on human-centric long-form videos rather than trying to simulate the entire physical universe, they are aiming directly at practical enterprise budgets: corporate training, explainer videos, and marketing demos.

For creators and small businesses, the tool is not quite ready for casual, on-the-fly use. Waiting 15 minutes for half a minute of low-res video means iteration cycles will be painful, though the actor royalty model is a surprisingly ethical touch that enterprises will love.

For the AI landscape, this is a loud reality check against the brute-force compute narrative. OpenAI raised over $6 billion in its latest round, yet a lean team with $2 million just blew past its duration limits by relying on clean proprietary data and clever parallel architecture. If small labs can repeatedly out-architect the giants on specific enterprise tasks, the narrative around AI moats is going to shift very quickly.