Movie Gen Video: State-of-the-art Video Generation
summary
Movie Gen is a text‑to‑video foundation model that generates HD videos up to 16 seconds by training a TA encoder plus a Transformer on hundreds of millions of videos and billions of images, using a 30B-parameter Transformer and 6,000 GPUs. It trains with flow matching instead of diffusion, compresses latent space by 8× (claimed >500× compute savings), and outputs smooth, higher‑resolution clips where a 10s HD sample equals ~500 million pixels. Benchmarks and human A/B tests show clear wins over prior systems.
AI-written summary · about
chapters (11)
- 0:00Movie Gen Video intro
- 1:28Text-to-video applications
- 3:17Text-to-video challenges
- 6:31Movie Gen foundation models
- 10:06Training improvements
- 13:59Scaling parameters and compute
- 16:29Multi-resolution pipeline
- 17:32Better training data
- 19:16Movie Gen Bench
- 20:20Scaling laws and future
- 21:22Conclusion and demo
related talks
33mHow Genmo Scales Video Generation with
26mDeploy Your Own VideoGen Model
17mThe State of Generative Media
24mState of Open Video Generation Models
14mEfficient Training of Video Generation Foundation Model at ByteDance
45mGenerative AI for Image and Video Production
34mHow NVIDIA is Advancing Video Curation with Generative AI
17mMorphe: High-Fidelity Generative Video Streaming with Vision Foundation Model