@thatguybg: Sora's video quality seems impossible so I dug into how it works under the hood

Brett Goldstein explains that Sora's impressive video quality is achieved using a combination of diffusion models and transformer architectures. Sora breaks videos into spacetime patches, which represent small fragments of space and time, and uses internal knowledge graphs to ensure coherent generation based on objects, actions, locations, and styles. Diffusion models refine these noisy patches into detailed frames, while transformers maintain temporal consistency across the video sequence. Although advanced, the technology still struggles with accurately replicating certain physical interactions.