Video Generation via Sequential Neural Network Upscaling

The patent divides the video generation task into multiple sequential neural networks, each responsible for generating video at a specific resolution stage. The initial network generates low-resolution video, and subsequent networks progressively upscale the resolution. This segmentation allows each network to focus on a specific resolution range, reducing the computational burden and training difficulty compared to a single network attempting to generate high-resolution video directly.

The patent employs classifier-free guidance during the generation process to preliminarily guide the neural networks toward temporally coherent outputs. By providing guidance signals that indicate desired temporal consistency properties, the networks can adjust their generation process in real-time to maintain coherence without requiring excessive computational resources for post-processing or iterative refinement.

The patent introduces classifier-free guidance as an intermediary mechanism that mediates between the training data and the neural network. Instead of directly optimizing the network to produce perfectly coherent videos, the classifier-free guidance acts as a mediator that provides incremental feedback signals during training, guiding the network toward temporally coherent outputs. This intermediary approach improves training efficiency by breaking down the complex optimization problem into smaller, more manageable steps.