Meta finally showed its hand against Sora, and Movie Gen is more than a video generator — it's a cast of foundation models that produces 1080p HD video across multiple aspect ratios with synchronized audio, supports precise instruction-based video editing, and can generate personalized videos based on a user's photo. This detailed review puts Meta's release head to head with OpenAI's Sora, breaking down where each system's architectural bets pay off, where the tradeoffs land, and how the two labs seem to be diverging on what a video foundation model should even look like.

The centerpiece is a 30B-parameter transformer trained with a maximum context length of 73K video tokens — enough for 16 seconds of generated video at 16 fps — with new SOTA claims across text-to-video synthesis, video personalization, video editing, video-to-audio, and text-to-audio generation. The review walks through the technical innovations Meta highlights: architecture and latent-space design, training objectives and recipes, data curation, evaluation protocols, parallelization techniques, and inference optimizations. If you're building or benchmarking generative video and audio systems, this comparison is a good compass for where the frontier is actually settling right now.