|
Join us virtually this Wednesday at 6pm to continue our monthly Paper Review series! This month we are looking at the concepts behind recent advancements in video generation. We will focus on Diffusion Transformers and Flow Matching techniques, culminating in a review of the recent Wan 2.1 technical report.
This virtual session will follow our usual format. In the first half, we will cover foundational concepts essential for understanding the paper and related techniques:
- Variational Autoencoders (VAEs)
- What is a latent space?
- Core principles of Diffusion Models
- Diffusion Transformers (DiTs) architecture
- Rectified Flow / Flow Matching concepts
The second half will then review the Wan paper, analyzing its components and contributions in the context of the concepts discussed:
- The overall model architecture and innovations
- Implementation and role of the Spatio-Temporal VAE
- Dataset construction and usage
- Details of the Captioning Pipeline for text-to-video/image
- Video Editing with Wan (VACE)
- Experiments in Video to Audio generation
|
