SkyReels-V3: Chinese AI model puts Elon Musk to work

While Western giants like OpenAI keep their best work under lock and key, Kunlun Wanwei in China has decided to go all in. The company has released SkyReels-V3, a video generation model aimed squarely at the heart of the short-form content industry, as open source. If you've ever tried to make an AI video, you know the main pain point: the character in the first frame rarely looks like the same person in the second.

SkyReels-V3 promises to end this facial "drift" and make video production genuinely predictable. As their trump card, the developers chose a face that's impossible to ignore. The digital clone of Elon Musk in their demo videos looks unnervingly stable.

He doesn't just stand there like a mannequin: he moves, gestures, and even tries his hand at livestream selling. Through it all, his face doesn't dissolve into a mush of pixels or morph into someone else's after five seconds. In an industry where every stray artifact means hours of post-production, this sounds like a death sentence for traditional production methods.

Under the hood, SkyReels-V3 relies on a new consistency control architecture. The model understands the geometry of a character's face and body and holds it in memory throughout the entire generation process. That lets it produce coherent, sequential scenes rather than disconnected clips.

Building a virtual influencer who runs a blog 24/7, with no days off and no lunch breaks, now becomes a technically trivial task. Kunlun Wanwei is clearly targeting the TikTok and Reels market, where a consistent persona is the key to brand recognition. Why does this matter right now?

We're watching the era of "hallucinogenic" video come to an end. Neural networks used to produce beautiful but uncontrollable dreams. Now we're getting tools with precise control.

The fact that SkyReels-V3 went open source rewrites the rules. Startups no longer need to spend millions of dollars training their own model from scratch: they can take a ready-made base and fine-tune it for specific needs, whether that's a sneaker ad or a full-length animated film. Of course, using Musk's likeness is both subtle trolling and a show of strength.

The developers are effectively hinting: "We can bring anyone to life, and we don't need permission to do it." That raises a huge range of ethical questions, but technological progress rarely pays attention to such trifles. While lawyers argue over the rights to digital faces, thousands of creators will already be churning out content with SkyReels-V3.

Ultimately, Kunlun Wanwei reinforces China's status as a leader in applied AI. They aren't just building something "like Sora"; they're making it accessible and usable for business here and now. Where a quality deepfake once required the skills of a professional video editor, today a mid-range GPU and a couple of lines of text are enough.

The barrier to entry into video production hasn't just dropped — it has vanished. The bottom line: SkyReels-V3 makes character consistency a technology available to everyone. Will paid services like Runway be able to justify their subscriptions when free alternatives let you create entire series with characters that never change?