Early AI-generated video models, like OpenAI’s original Sora, exhibited notable flaws, often glitching after just 15 seconds. However, advancements in technology have led to models such as Runway’s Gen-4 and Google’s Veo 3, which can create videos nearly indistinguishable from reality. This realism is attributed to their predictive capabilities regarding physical movements, light reflections, and water behaviors. Although these AI models operate as “black boxes,” making their inner workings opaque, they emulate reality through vast datasets of video inputs. Unlike traditional physics models used in engineering, which require extensive computation for precise predictions, neural networks offer a more efficient approach to approximate physical behaviors. Companies like Runway test their models by generating simple video scenarios and comparing them to real-life outcomes. Despite current imperfections, there is significant untapped potential in these models. Experts believe that incorporating more diverse data could lead to even more sophisticated world simulations.
Source link
Share
Read more