Phenaki

Phenaki is Google Research's contribution to long-form video generation, specifically addressing the challenge of generating temporally consistent video over extended durations from natural language narrative descriptions. Where most video generation models produce short clips of a few seconds, Phenaki explores generation of longer video sequences that follow a narrative arc.

Phenaki's architecture uses a token-based approach to video representation, compressing video into discrete tokens that can be generated autoregressively by a language model. This approach, inspired by text generation, allows the model to generate video sequences of variable length by predicting subsequent video tokens based on previous content and the text narrative.

The narrative conditioning allows Phenaki to generate videos that tell stories — changing scenes, evolving action, and sequential events described in a detailed text prompt. This narrative understanding distinguishes Phenaki from models that generate single static scenes rather than dynamic story sequences.

Phenaki demonstrates that the same autoregressive techniques successful in text generation can be applied to video generation with appropriate tokenization, contributing to the theoretical understanding of video generation approaches beyond diffusion models.

As a research model, Phenaki's primary impact is in demonstrating technical possibilities and contributing to the academic community's understanding of long-form video generation rather than providing a practical production tool.