视觉任务相对语言任务种类较多(detection, grounding, etc.)、粒度不同 (object-level, patch-level, pixel-level, etc.)，且部分任务差异较大，利用Tokenizer核心则为如何把其他模态映射到language space，并能让语言模型更好理解不同的视觉任务，更好适配LM建模方式，目前SOTA工作MAGVIT-v2，VideoPo

Paper name MAGVIT: Masked Generative Video Transformer Paper Reading Note Paper URL: https://arxiv.org/abs/2212.05199 Project URL: https://magvit.cs.cmu.edu/ Code URL: https://github.com/google-r

Paper name MAGVIT: Masked Generative Video Transformer Paper Reading Note Paper URL: https://arxiv.org/abs/2212.05199 Project URL: https://magvit.cs.cmu.edu/ Code URL: https://github.com/google-r