AI Trained Exclusively on Public Domain Films Generates New Videos
Video-generating artificial intelligence has been a hotly debated topic in the tech industry. These complex algorithms compile vast amounts of data to create short videos, often sourced from copyrighted material, sparking ethical and legal concerns, as well as questions about the energy required to produce AI-generated videos. Some tech giants argue that copyrighted content is a necessary input for AI development. However, a group of AI researchers at Moonvalley is challenging this notion.
Moonvalley, a Los Angeles-based AI startup, has developed a unique “3D-aware” video synthesis model, claiming that it is exclusively trained on public domain films. Their product, Marey, was initially released in March and is now available to the public via a credit-based system typical of AI video software. Notably, Moonvalley has garnered attention from industry professionals like Ed Ulbrich, a renowned VFX artist and producer known for his work on blockbuster films such as “Titanic” and “Benjamin Button.”
Ulbrich, who joined Moonvalley as a liaison to film studios, previously held reservations about generative AI but was impressed by Moonvalley’s approach. He praised the company’s commitment to ethical practices, emphasizing the absence of “stolen pixels” or content scraped from the internet. This ethical training model, according to Ulbrich, marks a significant shift in the AI landscape. Similarly, other AI initiatives have demonstrated success in training models solely on openly licensed or public domain data, challenging the narrative that copyrighted content is a prerequisite for AI advancement.
In a recent project, a team of AI researchers developed a large language model (LLM) using publicly available data, proving that extensive collections of copyrighted material are not essential for creating AI chatbots. Despite the arduous task of sifting through vast amounts of data – equivalent to 1,685,461 Bibles in this case – the resultant LLM rivaled established models like Meta’s Llama 1 and 2 7B. While the authenticity of Moonvalley’s public domain training data remains to be verified, their approach could challenge prevalent beliefs about data scarcity in the tech industry.
Moonvalley’s pioneering efforts to create AI models trained solely on public domain content may offer a compelling alternative to the prevailing trend of using copyrighted material. By prioritizing ethical sourcing and transparency, Moonvalley aims to reshape the narrative around AI development and promote a more sustainable approach to innovation in the field of video synthesis. As more companies and researchers explore the possibilities of generative AI without relying on proprietary data, the future of AI technology may become more inclusive, transparent, and ethically sound.