Researcher Insights on Sora ft. Google VideoPoet Author Lijun Yu and Nvidia Ethan He
This video is condensed from a one-hour live stream in Silicon Valley on March 16th, hosted by EntreConnect and IntelliGen.
Guests
Ethan He: Senior researcher engineer at NVIDIA working on open-source training frameworks of large models. He graduated from Carnegie Mellon University, worked at Facebook AI research, and has 6,000 academic citations for his papers.
Lijun Yu: Graduating PhD from Carnegie Mellon University, co-author of Google's text-to-video model, VideoPoet.
Chapters
0:00 Self-intro
1:01 Why is Sora powerful
2:35 Diffusion Transformer
7:27 Sora's limitations
10:40 Controversy on Training Data
11:58 What's Good Training Data
13:23 Human Feedback Loop
14:49 Training Infra
16:59 Model Size & Inference Time
18:20 Is Sora a world model
22:00 E2E vs. physics engine
23:02 Is computer vision solved
23:43 Implications to other fields
Disclaimer: Views are guests' own and do not represent that of their employers or Cindy's YouTube channel.
#OpenAI #ai #sora