Training-Free Efficient Video Generation via Dynamic Token Carving
Yuechen Zhang, Jinbo Xing, Bin Xia, Shaoteng Liu, Bohao Peng, Xin Tao, Pengfei Wan, Eric Lo, Jiaya Jia
2025-05-23
Summary
This paper talks about Jenga, a new method that lets AI models create high-quality videos much faster without needing extra training.
What's the problem?
Making videos with AI usually takes a lot of time and computer power, which slows down the process and makes it harder for people to use these tools for quick or creative projects.
What's the solution?
The researchers invented a technique called dynamic token carving and combined it with a way to build up video quality step by step, so the AI can focus on the most important parts of each frame and generate videos more quickly without losing quality.
Why it matters?
This matters because it makes video generation with AI much more practical and accessible, allowing more people to create videos faster for things like entertainment, education, or social media.
Abstract
Jenga, a novel inference pipeline for video Diffusion Transformer models, combines dynamic attention carving and progressive resolution generation to significantly speed up video generation while maintaining high quality.