NTH

dNaViT: Discrete Codes are General-Purpose Visual Representations

AuthorsXiaotong Li, Zhengrong Yue, Xu Huang, Jialv Zou, Minhao Jing, Wenhao Yan, Haodong Li, Hai Huang, Bo Zhao, Yan Feng, Peng Pei, Xunliang Cai

November 16, 2026 2 min read
Watch on YouTube
The one-line take

dNaViT turns images into reusable discrete codes designed to work across visual understanding, generation, and editing.

Key results

28.12
ImageNet-1K PSNR at 512×512

dB for dNaViT-G64 reconstruction.

885
OCRBench

Score for dNaViT-8B visual understanding.

49.5
GenEval2

Overall score for dNaViT-8B image generation.

52.17
Qwen-Image-Bench

Score for dNaViT-8B, above the reported open-weight baselines.

7.30
GEdit-Bench overall

Image-editing score for dNaViT-8B.

What the paper found

dNaViT asks whether a single discrete image vocabulary can preserve both semantic meaning and pixel-level detail. It keeps a pretrained Qwen3-VL vision encoder frozen, enriches its features with a zero-initialized residual modulator, then applies grouped product quantization to expand code capacity without lengthening the spatial token grid. A staged dual-branch objective trains the same codes to reconstruct semantic features and image pixels; a code-conditioned flow refiner sharpens decoded images. At a spatial downsampling factor of 32, dNaViT-G64 reaches 28.12 dB PSNR and 0.81 SSIM at 512×512 on ImageNet-1K. Paired with Qwen3-8B, it scores 885 on OCRBench and 96.2 on DocVQA, putting it near Qwen3-VL-8B on document understanding. For image generation, dNaViT-8B scores 49.5 on GenEval2 and 52.17 on Qwen-Image-Bench, exceeding open-weight baselines on the latter while trailing OpenAI’s GPT Image 1. On GEdit-Bench, it earns an overall score of 7.30. Across 12 visual-understanding benchmarks, 4 generation benchmarks, and an editing benchmark, the results support using one discrete representation across reconstruction, understanding, generation, and editing.

Original abstract

Discrete autoregressive modeling, notably through next-token prediction, has become a cornerstone of modern language models. Extending this paradigm to vision allows visual and language models to converge on a unified framework, thereby fostering shared methodology and infrastructure. This work investigates a central question: how far can a finite discrete vocabulary go in representing the richness of the visual world? We present dNaViT, a Discrete Native-resolution Vision Transformer that produces discrete visual primitives analogous to language tokens. Built upon a frozen pretrained encoder, dNaViT applies lightweight residual modulation with a dual-branch objective to inject reconstruction-aware cues into semantic features. By leveraging grouped product quantization, it scales discrete capacity without inflating spatial sequence lengths, while a balanced training recipe maintains both semantic and pixel-level information. Consequently, dNaViT functions as a versatile discrete visual tokenizer, enabling unified encoding across downstream tasks alongside high-fidelity image decoding. Extensive evaluations show that dNaViT matches or exceeds task-specific baselines across 12 visual understanding benchmarks, 4 generation benchmarks, and an image editing benchmark. These findings demonstrate that a compact discrete vocabulary can serve as a universal visual currency and unlock the full potential of fully unified, native multimodal methodology.

Read the original paper

More in Generative Models

Browse all 87 papers →