File size: 428 Bytes
881a7e3
1654d6c
 
 
 
 
 
 
 
 
 
 
881a7e3
1
2
3
4
5
6
7
8
9
10
11
12
13
## MTVBenth: Benchmarking Video Generation Models with Multiple Transition Text Prompts

Submission to NeurIPS 2025 Dataset and Benchmark track.


Xiaodong Cun, Xiuli Bi, Ruihuan Yang, Jianfei Yuan, Bin Xiao, Bo Liu.

GVC Lab, Great Bay University, and Chongqing University of Posts and Telecommunications


We list all the generated videos here in each folder.

The prompt benchmark we collected is shown in the `prompt` folder