How Genmo Scales Video Generation with
summary
Genmo claims production text-to-video at scale: over 1.5 million users, thousands of GPUs across 20+ data centers, and a training supercomputer ranked top‑50 and 98% carbon free. Engineering choices include diffusion models for generation, Ray Data and a custom GFE edge layer to get first pixels in ~10 seconds, and a predictive autoscaler plus aggressive health checks to tolerate 2–5% GPU faults and spot preemptions. Result: 5-second clips take ~1 minute to render while running ~11x cheaper than managed alternatives.
AI-written summary · about
chapters (18)
- 0:00Intro
- 0:32Our Highle Mission
- 2:25What is Genmo
- 3:20Scaling Genmo
- 5:43Early Days
- 7:53Global Growth
- 9:01Control Plane Issues
- 11:20Data Plan Fault Tolerance
- 13:45Geo Distributed User Base
- 17:57Ice
- 19:18Models
- 21:34Longterm impact
- 23:07Controllability
- 23:51Diffusion Models
- 24:58Pipeline Latency
- 26:32Heterogenous GPUs
- 28:51Train your data differently
- 31:18More scientific approach
related talks
22mMovie Gen Video: State-of-the-art Video Generation
26mDeploy Your Own VideoGen Model
17mThe State of Generative Media
14mEfficient Training of Video Generation Foundation Model at ByteDance
24mState of Open Video Generation Models
17mMorphe: High-Fidelity Generative Video Streaming with Vision Foundation Model
20mBatchGen: An Architecture for Scalable and Efficient Batch Inference
45mGenerative AI for Image and Video Production