A China-based startup, Sand AI, has launched an overtly licensed video-generating AI mannequin that’s garnered reward from entrepreneurs like Microsoft Analysis Asia founding director Kai-Fu Lee. However Sand AI seems to be censoring pictures which may elevate the ire of Chinese language regulators from the hosted model of the mannequin, in keeping with TechCrunch’s testing.
Earlier this week, Sand AI introduced Magi-1, a mannequin that generates movies by “autoregressively” predicting sequences of frames. The corporate claims the mannequin can generate high-quality, controllable footage that captures physics extra precisely than rival open fashions.
👀 You gained’t consider that is AI-generated
🧠 You gained’t consider it’s open-source
🎬 You gained’t consider it’s FREEMagi-1 video mannequin simply humiliated industrial video instruments
Particulars and examples beneath: 👇 pic.twitter.com/zlXRecWeqH
— Farhan (@mhdfaran) April 21, 2025
Magi-1 is simply too impractical to run on most client {hardware}. It’s 24 billion parameters in measurement, and requires between 4 and eight Nvidia H100 GPUs to run. (Parameters are the inner variables fashions use to make predictions.) For a lot of customers — this reporter included — Sand AI’s platform is the one place they’ll take a look at drive Magi-1.
The platform wants a “immediate” picture to kick off video era. Not all prompts are permissible, TechCrunch shortly found. Sand AI blocks picture uploads of Xi Jinping, Tiananmen Sq. and Tank Man, the Taiwanese flag, and insignias supporting Hong Kong liberation. Filtering seems to be occurring on the picture degree; renaming picture recordsdata didn’t skirt the blocking.
Sand AI isn’t the one Chinese language startup stopping uploads of politically delicate pictures to its video era software. Hailuo AI, Shanghai-based MiniMax’s generative media platform, blocks images of Xi Jinping as properly. However Sand AI’s filtering seems to be significantly aggressive; Hauiluo permits pictures of Tiananmen Sq..
As Wired defined in a bit from January, fashions in China are required to comply with stringent data controls. A 2023 regulation forbids fashions from producing content material that “damages the unity of the nation and social concord” — that’s, counters the federal government’s historic and political narratives. To conform, Chinese language startups usually censor their fashions, both by prompt-level filters or fine-tuning.
Curiously, whereas Chinese language fashions have a tendency to dam political speech, they usually have fewer filters than their American counterparts for pornographic content material. 404 recently reported that a lot of video mills launched by Chinese language corporations lack fundamental guardrails that stop individuals from producing nonconsensual nudity.