先说结论。它和通用视觉自监督的关系在于：用 learnable global merging 做可变长视觉 tokenization，和视觉 latent 压缩/对齐有关。 中相关；详见方法、贡献和实验边界。
Figureure 1 · CKNNA and gFID measurements from the nested dropout tokenizer and the Figure 1. CKNNA and gFID measurements from the nested dropout tokenizer and the diffusion model trained on it, respectively. Upper: CKNNA between untruncated $( N = 2 5 6 )$ and truncated $( M < 2 5 6 )$ latents; the black dotted line shows CKNNA for an independently trained tokenizer (vs. 256). Lower: gFID of the variable-length diffusion model learned jointly across token lengths vs. separately for each length.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 Variable-Length Tokenization via Learnable Global 的方法或实验，请结合正文精读段落一起看。Figureure 2 · Overview of our methodFigure 2. Overview of our method. To encourage representational alignment, we propose length modulation via merging (Sec. 3.1). The merging operation is image-agnostic (i.e., globally determined) but optimized for tokenization and alignment (Sec. 3.2).这张图概括 Variable-Length Tokenization via Learnable Global 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。
核心问题
它和通用视觉自监督的关系在于：用 learnable global merging 做可变长视觉 tokenization，和视觉 latent 压缩/对齐有关。
方法拆解
用 learnable global merging 做可变长视觉 tokenization，和视觉 latent 压缩/对齐有关
