Meituan's dNaViT quantizes images into LLM-style discrete tokens

A discrete vocabulary of image tokens, used the same way language models use word pieces, can match specialized continuous representations across "12 visual understanding benchmarks, 4 generation benchmarks, and an image editing benchmark," according to a new paper on Hugging Face from Meituan's LongCat team.

The system is called dNaViT, short for Discrete Native-resolution Vision Transformer. It keeps a pretrained Qwen3-VL ViT frozen, bolts on a lightweight residual modulator, and runs the output through grouped product quantization, which scales code capacity through combinations rather than one monolithic codebook. At 512×512 on ImageNet-1K the G64 variant reports 28.12 dB PSNR; at 1024×1024 it reports 32.90 dB. On the linear-probe ImageNet-1K task the discrete tokens land at 79.47%, inside a point of the continuous Qwen ViT baseline at 80.45%.

The authors frame the work as a bet on representation, not a benchmark win. "This work demonstrates that a shared discrete token space can serve as a universal visual currency, providing a general-purpose intermediate representation across complementary tasks and a foundation for fully unified, native multimodal modeling," the paper says.

Plugged into an 8B language-model stack, dNaViT scores 70.7 on MMMU and 49.5 on GenEval2. Weights ship through Hugging Face under the LongCat-Next repo.

The paper flags gaps too. OCRBench trails Qwen3-VL-8B at 885 versus 896, and small-text reconstruction lags stronger baselines like FLUX.1. Computer-vision releases have been a steady drumbeat on our tracker, with 62 stories in the last 90 days.