LongCat team (Xiaotong Li et al.), arXiv · Discrete Visual Tokenizer

dNaViT: Discrete Codes are General-Purpose Visual Representations

dNaViT is a discrete native-resolution vision transformer that produces discrete visual tokens analogous to language tokens. It is built on a frozen pretrained encoder, uses lightweight residual modulation with a dual-branch objective, and applies grouped product quantisation to raise discrete capacity without lengthening spatial sequences. The authors report that it matches or exceeds task-specific baselines across 12 visual understanding benchmarks, 4 generation benchmarks and an image editing benchmark, and that it also supports high-fidelity image decoding.