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.
Figures & tables
Figure 1: dNaViT: A unified visual tokenizer for general-purpose modeling. dNaViT represents arbitrary-resolution images as discrete visual primitives, natively unifying reconstruction, understanding, generation, and editing under a shared vocabulary. By achieving performance competitive with specialized baselines across all domains, dNaViT validates a shared token-based paradigm for vision, analogous to language modeling.
Figure 2: dNaViT’s unified discrete representation as a foundation for diverse downstream visual tasks. Qualitative showcases illustrate the versatility of this shared representation.
Figure 3: Overview of dNaViT. (a) A frozen representation encoder handles arbitrary resolutions, with ResMod injecting multi-level information before group-wise product quantization. The shared discrete codes are decoded by individual ViT decoders into pixel and semantic reconstructions. (b) ResMod fuses intermediate features and predicts a zero-initialized scale-and-shift residual. (c) Pixel and semantic branches are optimized with image and feature reconstruction losses, respectively.
Figure 4: Tokenization and Unified Codes for Downstream Tasks. (a) At spatial position t , the modulated feature zt is split into G channel groups and quantized with group-specific codebooks E(g) , yielding indices (kt(1),…,kt(G)) and their embeddings. (b) Understanding supplies only code indices to the language model (LM) through newly initialized group-wise embedding tables {W(g)}g=1G , trained jointly with the LM. (c) Reconstruction concatenates the indexed embeddings from the frozen tokenizer codebooks and feeds them to the pixel decoder Dpix . (d) Generation and editing use a multi-sub-token head atop the LM to predict each position’s multi-level indices, optionally conditioned on a reference image, then decode through the same frozen codebooks and pixel decoder.
Figure 5: Joint preservation of high-level semantics and low-level pixel detail in a shared discrete representation. (a) dNaViT’s reconstructed semantic features closely match the original QwenViT features. Each cell reports the cosine similarity between corresponding tokens. (b) Pixel reconstructions from dNaViT-G16 and dNaViT-G64, with matched crops highlighting the fine-grained text on the jersey.
Figure 6: Overview of the code-conditioned Refiner. Code embeddings condition a flow decoder through global context tokens and time-conditioned spatial residuals. A multi-step teacher is distilled into an efficient one-step student with adversarial supervision. The timestep t in the diagram corresponds to τ in the equations. Increasing the quantization group count (G64) improves reconstruction fidelity, whereas combining fewer groups (G16) with the one-step Refiner preserves fidelity while enhancing perceptual quality.
Figure 7: Three-stage training schedule of dNaViT. Continuous pixel pretraining is followed by discrete semantic reconstruction and joint reconstruction. Stage 2 stops pixel gradients at the quantized embeddings. Stage 3 opens this gradient path while a frozen copy of the Stage-2 ResMod supplies the feature target. Flame and snowflake icons mark trainable and frozen modules; dashed purple paths denote fixed targets.
Method
Configuration
ImageNet
Resolution
f
Spatial positions
PSNR (dB) ↑
SSIM ↑
Continuous representations
SD-VAE 1.x / 2.x ( Rombach et al., 2022 )
2562
8
1024
23.54
0.68
OmniTokenizer ( Wang et al., 2024a )
2562
8
1024
26.74
0.82
SD-VAE XL ( Podell et al., 2023 )
2562
8
1024
27.37
0.78
Qwen-Image ( Wu et al., 2025a )
2562
8
1024
32.18
0.90
Table 1: ImageNet reconstruction across resolutions. f denotes the spatial downsampling factor; spatial positions exclude quantization groups and do not specify bit rate. All dNaViT results use the same evaluation protocol on the full set of 50,000 ImageNet validation images. Higher is better for both metrics. Baseline results follow source-specific protocols and are not directly comparable across sources.
Method
f
Resolution@ 512×512
Resolution@ 1024×1024
T-ACC ↑
T-NED ↑
F-Sim ↑
T-ACC ↑
T-NED ↑
F-Sim ↑
Continuous Autoencoders
DC-AE ( Chen et al., 2024a )
32
38.25
56.22
0.39
52.14
68.02
0.50
VA-VAE ( Yao and Wang, 2025 )
16
53.30
69.65
0.57
62.84
77.21
0.69
SD-XL ( Podell et al., 2023 )
8
56.86
72.55
0.55
67.84
81.15
0.70
SD-3.5 ( Stability AI, 2024 )
8
81.84
90.30
0.81
89.71
95.05
0.93
Table 2: Fine-grained reconstruction on TokBench. Results are macro-averages over three instance-scale buckets. T-ACC and T-NED are percentages; F-Sim is cosine similarity. f is the spatial downsampling factor.
Method
Resolution
T-ACC ↑
T-NED ↑
F-Sim ↑
S
M
L
S
M
L
S
M
L
dNaViT-G16
5122
3.52
24.56
72.66
17.60
52.52
87.19
0.14
0.25
0.61
dNaViT-G64
5122
12.18
54.94
87.95
34.73
76.67
94.76
0.29
0.49
0.84
dNaViT-G16
10242
10.89
38.81
85.44
28.93
65.58
93.74
0.22
0.35
0.64
dNaViT-G64
10242
26.79
68.50
93.95
51.93
85.29
97.58
0.49
0.65
0.86
Table 3: Scale-stratified reconstruction with dNaViT on TokBench. S, M, and L denote the official small, medium, and large instance buckets. Text metrics are percentages. Buckets are resolution-specific and do not define a fixed-instance comparison across resolutions.
Method
MMMU
MMMU-Pro
MathVista
MathVision
RealWorldQA
MMStar
Continuous
LLaVA-OV-1.5-8B ( An et al., 2025 )
55.4
31.3
69.6
25.6
68.1
67.7
LLaVA-OV-2-8B ( An et al., 2026b )
–
–
–
–
69.7
64.8
InternVL3-8B ( Zhu et al., 2025 )
62.7
–
71.6
29.3
70.8
68.2
InternVL3.5-8B ( Wang et al., 2025b )
73.4
–
78.4
56.8
67.5
69.3
Kimi-VL-A3B ( Kimi Team, 2025 )
57.0
36.0
68.7
21.4
68.1
61.3
Table 4: STEM reasoning and general multimodal understanding. Comprehensive comparison of vision-language models (VLMs) using continuous or discrete visual representations.
Method
f
Latent Shape
IN-1K
IN-A
IN-R
IN-V2
IN-Sketch
ObjectNet
Continuous
QwenViT ( Bai et al., 2025a )
32
16×16
80.45
60.13
65.00
70.18
46.57
56.48
InternViT ( Wang et al., 2025b )
28
16×16
76.19
30.23
59.98
65.68
35.77
46.36
MoonViT ( Kimi Team, 2025 )
28
16×16
75.40
23.96
49.24
63.34
34.88
45.46
dNaViT-G16 †
32
16×16
80.40
60.25
65.19
70.18
46.56
56.48
Discrete
Table 5: Image classification on ImageNet-1K and distribution-shift benchmarks. Frozen representations are evaluated by linear probing on ImageNet-1K and direct transfer to ImageNet-A, ImageNet-R, ImageNet-V2, ImageNet-Sketch, and ObjectNet. All results report top-1 accuracy (%). † denotes continuous dNaViT features before quantization.
Method
OCRBench
OCRBench V2 (en/zh)
CharXiv (DQ/RQ)
DocVQA
InfoVQA
OmniDocBench (en/zh) ↓
Continuous
LLaVA-OV-1.5-8B ( An et al., 2025 )
829
–
70.9/–
95.0
78.4
–
LLaVA-OV-2-8B ( An et al., 2026b )
782
–
–
95.2
74.4
–
InternVL3-8B ( Zhu et al., 2025 )
880
–
73.6/37.6
92.7
76.8
–
InternVL3.5-8B ( Wang et al., 2025b )
832
–
72.2/44.4
92.3
76.2
–
Kimi-VL-A3B ( Kimi Team, 2025 )
867
–
–
–
83.2
–
Table 6: OCR, document, and chart understanding. Comprehensive comparison of vision-language models (VLMs) using continuous or discrete visual representations.
Model
Comp.
Params.
GenEval2 GM
QwenImage Bench
Spatial GenEval
LongText EN
LongText CN
Proprietary Models
GPT Image 1 ( OpenAI, 2025 )
–
–
–
54.07
0.61
0.96
0.62
Seedream 4.0 ( ByteDance Seed, 2025 )
–
–
–
56.21
0.63
0.92
0.93
Seedream 4.5 ( ByteDance Seed, 2026 )
–
–
–
56.78
–
0.99
0.99
GPT Image 2 ( OpenAI, 2026 )
–
–
–
64.69
–
0.96
0.98
Nano Banana ( Google, 2025b )
–
–
44.6
–
0.62
0.87
0.33
Table 7: Text-to-image generation. GenEval2 GM is the GenEval2 overall score; QwenImageBench is the overall score on Qwen-Image-Bench (0–100); SpatialGenEval measures spatial-relation generation; LongText reports the English and Chinese subsets of LongText-Bench. Comp. is per-axis spatial compression, including latent packing.
Model
Comp.
Params.
SC ↑
PQ ↑
Overall ↑
Proprietary Models
Gemini 2.0 ( Kampf and Brichtova, 2025 )
–
–
6.73
6.61
6.32
Gemini 2.5 Flash Image ( Fortin et al., 2025 )
–
–
7.41
7.96
7.10
GPT-Image-1 [High] ( OpenAI, 2025 )
–
–
7.85
7.62
7.53
Open Models
AnyEdit ( Yu et al., 2024a )
8×
0.9B
3.18
5.82
3.21
Table 8: Image editing on GEdit-Bench (English full set). SC: semantic consistency; PQ: perceptual quality. Higher is better for all three metrics. Comp. is per-axis spatial compression, including latent packing.
Figure 8: Pixel and representation losses for residual feature fusion. Compared with SelfDistill, ResMod maintains lower pixel MSE and representation cosine losses during late-stage training, indicating better preservation of both visual details and semantic representations.
Exp ID
Type
ResMod
MMMU
RealWorldQA
MMStar
OCRBench
TextVQA
AI2D
ChartQA
DocVQA
Avg. ↑
1
Continuous
✗
39.44
60.26
44.60
695
74.84
81.87
73.64
86.29
66.31
2
✓
40.56
60.92
45.33
695
74.89
81.51
73.16
86.01
66.48
3
✓
40.00
62.61
46.60
713
75.05
81.31
72.80
86.01
66.96
4
Discrete
✓
38.33
59.08
45.73
682
73.49
80.60
71.00
83.98
65.05
Table 9: Visual understanding with continuous and discrete representations. Continuous features are evaluated with and without ResMod. OCRBench is divided by 10 when computing the eight-benchmark average.
Figure 9: Qualitative reconstruction comparison at the same input resolution. f denotes the spatial downsampling factor. Continuous and Discrete denote unquantized latent representations and quantized codes, respectively. G and R specify the number of quantization groups and refinement steps.
Figure 10: Full-page OCR of a densely printed page from The Illustrated London News (1846). The response transcribes the multi-column body text, headings and engraving captions in reading order.
Figure 11: Full-page OCR of a dense Chinese research article, covering the bilingual header, title, author affiliations with superscripts, abstract and body text.
Figure 12: Full transcription of a 1911 cursive letter written on printed hotel stationery, covering both the printed letterhead and the connected handwriting.
Figure 14: Mechanical-vibration reasoning from a schematic: the model identifies series, parallel and lever-arm spring combinations (MMMU, hard).