This paper reviews recent work that treats video generation models as “world models” and focuses on one central problem: these models can in principle simulate physics and long-term cause-and-effect, but they are currently too slow and costly for practical use. The authors organize the field around efficiency. They propose a three-part taxonomy that groups methods by efficient modeling paradigms, efficient network architectures, and efficient inference algorithms. They argue that closing the efficiency gap is essential for real-time, interactive uses such as autonomous driving, embodied robots, and game simulation.

What the authors did is a systematic, literature-based survey rather than a new experiment. They collect and summarize techniques across different video-generation approaches. These include diffusion-based methods (which produce video by iteratively “denoising” random input), flow-matching methods (which view generation as moving samples along a continuous path), and auto-regressive models (which predict frames step by step). On the architecture side they review designs like variational autoencoders (VAE, a way to compress images into a smaller internal representation), hierarchical models, memory mechanisms for long context, and attention methods. For inference they cover practical tricks such as distillation (making a slow model faster), parallelism, caching, pruning, and quantization (making models use less memory and compute).

At a high level, the efficiency ideas reduce the work the model must do. Working in a compact latent space instead of full pixels cuts the data size. Memory mechanisms and long-context strategies let a model keep important history without storing every frame. Efficient attention and other architectural choices reduce the number of expensive operations. For sampling-heavy methods like diffusion models, distillation and improved samplers aim to cut the many iterative steps that cause latency. For stepwise auto-regressive models, careful handling of key-value caches helps avoid explosive memory use as sequences grow.