Reproduce “dNaViT: Discrete Codes are General-Purpose Visual Representations” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xiaotong Li, Zhengrong Yue, Xu Huang, Jialv Zou, Minhao Jing, Wenhao Yan, Haodong Li, Hai Huang and 4 more

dNaViT uses a frozen vision encoder with residual modulation and grouped product quantization to create discrete image codes trained for semantic and pixel reconstruction. On 512-by-512 ImageNet-1K validation images, dNaViT-G64 scores 28.12 dB PSNR and 0.81 SSIM. Paired with Qwen3-8B for multimodal training, dNaViT-8B scores 885 on OCRBench, below Qwen3-VL-8B’s 896 but above Kelix-8B’s 867.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.