关于Bagel视频生成范式的疑问

很棒的工作，通读论文发现在连续图像预测（视频生成训练）时，group一组将要生成的token并进行预测(full-attention)；但是观察代码，在生成图片的时候仍然是自回归的生成，并没有group解码(例如每个group size的token连续预测)，团队有没有测试过group扩散和单个自回归扩散谁的连续帧生成效果更好呢？

The issue names no files, tests, or entry points. Start by locating the video-generation inference path and comparing it with the paper’s grouped-token training description. Done would require a documented comparison of grouped decoding and single-token autoregressive diffusion for continuous-frame generation.

Written by the indexing model from the issue text.