Sand AI Launches Magi-1 Video Generation Model Amid Controversial Image Restrictions

Alongside the release of this transformative technology, Sand AI has open sourced its own video-generating AI model, called Magi-1. This state-of-the-art model produces videos by modeling sequences of frames frame-by-frame in an autoregressive way. This approach allows for the smooth integration of engaging, rich, interactive storytelling elements. Rated at a massive 24 billion parameters, Magi-1 is a historic breakthrough in video generation technology.

Magi-1 needs a lot of computational resources to work optimally. In fact, it needs between four and eight Nvidia H100 GPUs. The model’s extreme hardware requirements render it unlikely for nearly any consumer-level applications. This means that only deeply resourced organizations and institutions with their own advanced computing infrastructure are able to leverage it.

Nevertheless, the debut of Magi-1 animated a vigorous discussion about its limits on politically sensitive content. Sand AI has recently released a handful of new image upload blocks. These bans single out graphic depictions of Xi Jinping, Tiananmen Square, Tank Man, the Taiwanese flag and symbols promoting Hong Kong independence. These restrictions are in line with a 2023 law that prohibits models from generating content deemed to “damage the unity of the country and social harmony.” Such regulations are a manifestation of the Chinese government’s attempts to continue to suppress historical and political narratives from existing in digital spaces.

Sand AI’s censorship is strikingly similar to a broader pattern occurring in Chinese AI models. These models are inclined to more narrowly police political speech than their counterparts in other countries. American models tend to use much stricter filters for content considered pornography. In contrast, Chinese models show more flexibility in those aspects. This difference has broader implications on the interplay between freedom of expression and regulatory requirements across jurisdictions.

The impact of these muzzles is enormous. Together, they determine the limits of the Magi-1 model and wider ecosystem of AI-generated content in China. As we’ve heard from those analysts, these restrictions can kill creativity and innovation. Developers may be reluctant to produce politically sensitive content out of fear of future retaliation.