The authors propose Future Chunk Prediction (FCP), an auxiliary objective jointly optimized with standard teacher-forced denoising in autoregressive video diffusion.
Future chunk prediction improves physical plausibility in autoregressive video generation
Authors show that an auxiliary objective to predict temporally separated chunks, without seeing intervening frames, yields better physical plausibility and more structured representations.
Chinese Tech
Ruopeng Gao · Yukun Huang · Jiarui Hu · Shuai Wang · Wenbo Hu · Ying Shan · +1 more
Nanjing University · ARC Lab, Tencent IEG · Zhejiang University
Research Digest··2 min read
The authors introduce Future Chunk Prediction (FCP), an auxiliary objective for training autoregressive video diffusion models.
§