Seedance 1.0: New #1 Video Generator - Architecture, Data, Training, Science, Optimizations - Paper

Seedance 1.0: New #1 Video Generator - Architecture, Data, Training, Science, Optimizations - Paper

paper - arxiv.org/pdf/2506.09113

Code DeepSeek V3 From Scratch Full Course -    • Understand & Code DeepSeek V3 From Scratch...  

support me on patreon - www.patreon.com/vukrosic/membership

contact: vukrosic1@gmail.com

0:00 - Intro & Importance
0:27 - Model Overview
1:01 - VAE: Smart Compression
3:29 - Latents vs Pixels
5:22 - Compression Details
7:54 - VAE Training Losses
10:41 - Diffusion Transformer
12:16 - Spatial vs Temporal
14:51 - Text & Vision Attention
16:59 - 3D/1D Positional Info
20:02 - HD Video Upscaling
21:28 - Prompt Rewriting
24:20 - Dataset Curation
28:02 - Caption Generation
28:27 - Processing at Scale
29:07 - Multi-Stage Training
31:06 - Image to Video
31:47 - Specialist Merging
32:37 - RL with Human Feedback
34:16 - Training Optimizations
36:50 - Fast Inference Tricks
38:03 - Final Thoughts