Autoregressive image generation has recently emerged as a competitive visual generation paradigm, yet its generation order is often predefined or heuristic, such as next-patch prediction, random-order prediction, or one-step generation for specific scales. We observe that these strategies can cause models to overlook important structural dependencies in images, thereby limiting both structural robustness and reasoning capability. In this work, we propose a context-aware autoregressive generation framework, where the model dynamically determines the next token location based on global context and current local information. Rather than following a fixed or random order, our method allows the generation process to adapt to image structure and semantic dependencies. Extensive experiments show that context-aware generation significantly improves structural robustness across different image AR paradigms, including masked autoregressive and scale-wise autoregressive models. Beyond quantitative improvements, we further observe emerging reasoning properties: the model can better understand complex semantic instructions and generate images that more faithfully follow the given guidance. These findings suggest that generation order is a critical yet underexplored factor in image autoregressive modeling, and that context-aware decoding can unlock stronger reasoning-aware visual generation capabilities. Code is available at https://github.com/Hpesojyjx/Saccades-AR.
Figures & tables
Figure 1: Generation order shapes reasoning-like behaviors in autoregressive image generation. With random-order decoding (top), intermediate predictions remain scattered and incomplete, weakening feedback for counting (e.g., “four vases") and often causing over-generation. With context-aware location selection (bottom), the same model and seed tend to complete objects more coherently (roughly one-by-one) and stop adding new ones once the count is met.
Figure 2: Comparison to one-step and random-order generation paradigms. Our SAR model decides the generation order based on each token’s latent distance to a global token. This is only a minimal change for image AR models, but we find it crucial for the AR model’s reasoning behavior.
Figure 3: Infinity-8B: outline-to-detail reasoning under SAR. In each strip, the top row is random ordering and the bottom row is SAR; columns are AR steps. Within the first few SAR steps the global outline already encodes the correct count (four candles, six houseplants), the correct object set with no hallucinated extras (fox+rabbit), and the correct global geometry (circular pizza). Subsequent steps fill in textures without changing those structural decisions. Random ordering instead grows partial textures everywhere and frequently fails to recover count, identity, or shape by the final step.
Figure 4: MaskGen-XL: object-by-object reasoning under SAR. Top to bottom: “four vases”, “a child walking to the right of an adult down a quiet street”, and “a birthday cake with colorful candles on a kitchen counter”. SAR completes one object before starting the next (counting), commits one entity before introducing the other (spatial relation), and follows scene → main object → parts (compositional hierarchy); random ordering grows everything concurrently and only resolves identity and layout in the last few steps.
Model
Order
Cnt.
Spat.
Attr.
Exist.
Ord.
View
Avg.
MaskGen-XL
random
2.55
2.24
2.39
2.51
2.32
2.51
2.42
SAR
3.08
2.78
2.94
3.01
2.82
2.89
2.92
Infinity-2B
random
2.76
2.43
2.64
2.72
2.49
2.68
2.62
SAR
3.27
2.98
3.16
3.22
2.99
3.10
3.12
Infinity-8B
random
3.09
2.74
2.94
3.02
2.82
2.91
2.92
SAR
3.61
3.31
3.45
3.49
3.30
3.36
3.42
Table 1: Results on our complex long-prompt benchmark. Scores are judged by Gemini 2.5 Pro on a 1–5 scale.
Model
Method
S.Obj
T.Obj
Count
Colors
Pos.
C.Attr
Overall
MaskGen-L
Random
100.00
54.55
40.00
85.11
18.00
26.00
53.94
SAR
100.00
56.57
50.00
89.36
20.00
29.00
57.49
MaskGen-XL
Random
98.75
58.59
43.75
87.23
21.00
28.00
56.22
SAR
98.75
64.65
46.25
88.30
25.00
37.00
59.99
Infinity-2B
Random
100.00
82.62
64.38
82.77
42.10
58.88
71.79
SAR
100.00
83.34
65.27
85.42
41.00
60.38
72.57
Table 2: GenEval pass rate (%, ↑ ) for both AR paradigms. Best per (model, metric) pair in bold.
Model
Method
Entity
Attribute
Relation
Global
Other
Overall
MaskGen-L
Random
81.82
80.04
80.91
76.12
81.08
70.11
SAR
78.30
80.84
82.91
82.75
81.70
70.51
MaskGen-XL
Random
79.33
80.98
81.63
80.81
80.52
71.00
SAR
81.53
79.82
82.83
82.81
81.08
71.57
Infinity-2B
Random
89.14
86.19
92.50
83.18
82.67
82.71
SAR
89.24
86.23
92.51
82.88
82.93
82.76
Table 3: DPG-Bench L1 category breakdown (%, ↑ ) for both AR paradigms. Best per (model, metric) pair in bold.
Patch-wise (MaskGen)
Scale-wise (Infinity)
Order
MaskGen-L
MaskGen-XL
Infinity-2B
Infinity-8B
Random
6.79
6.75
10.867
9.332
SAR
6.79
6.74
10.970
9.317
Table 4: MJHQ-30K FID ( ↓ ) for both AR paradigms. Across all four scales, SAR is statistically tied with the random-order baseline, indicating that context-aware ordering does not change the distribution-level aesthetic quality of the generated images.
MaskGen-XL
Infinity-8B
Order
GenEval ( ↑ )
DPG-Bench ( ↑ )
GenEval ( ↑ )
DPG-Bench ( ↑ )
Random
0.55
71.00
0.804
86.05
Sequential
0.57
71.00
0.799
86.01
Inside-out
0.55
71.30
0.808
86.25
SAR (ours)
0.58
71.57
0.815
86.36
Table 5: Ablation on the inference-time generation order, on one representative model from each AR family. “Sequential” generates tokens one by one in their natural index order; “Inside-out” starts from the middle token and expands outward. Higher is better for all columns; best per column in bold.
Selection rule
FID ↓
Recall ↑
Confidence (original)
6.48
0.41
SAR
5.95
0.43
Table 6: Controlled MaskGIT-Base comparison: only the position-selection criterion changes.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Patch-wise (MAR)
Scale-wise (VAR)
Order
MAR-Base
MAR-Large
VAR- d16
VAR- d20
Random
2.29
1.78
3.56
2.77
SAR (class similarity)
2.21
1.75
3.48
2.73
Appendix
Table 7: Class-conditional FID ( ↓ ) on ImageNet 256×256 across both AR paradigms. Within each column, both rows share the same checkpoint (the VAR rows use a retrained checkpoint with M=4 multi-step inference at the last scale) and the same sampling configuration; only the inference-time generation order differs.
Model
Order
FID ↓
IS ↑
Precision ↑
Recall ↑
VAR- d16
1-step ∗
3.30
274.4
0.84
0.51
1-step
3.42
268.5
0.83
0.49
random
3.24
278.2
0.84
0.52
SAR
3.05
286.2
0.86
0.54
VAR- d20
1-step ∗
2.57
302.6
0.83
0.56
1-step
2.68
297.0
0.82
0.54
Appendix
Table 8: Results of retraining VAR models with different generation orders. For each VAR model, we apply the SAR order at the last scale during retraining. The superscript ∗ indicates the original setting where both training and evaluation generate all tokens within each scale in one step.
Model
Order
Time (s/image) ↓
FID ↓
IS ↑
Prec. ↑
Rec. ↑
VAR- d16
1-step
0.073±0.002
3.42
268.5
0.83
0.49
VAR- d16
random
0.093±0.002
3.24
278.2
0.84
0.52
VAR- d16
SAR
0.093±0.002
3.05
286.2
0.86
0.54
VAR- d20
1-step
0.088±0.002
2.68
297.0
0.82
0.54
VAR- d20
random
0.113±0.002
2.53
304.8
0.84
0.57
VAR- d20
SAR
0.113±0.002
2.35
314.5
0.85
0.59
Appendix
Table 9: VAR generation quality and end-to-end inference latency. Times are mean ± population standard deviation over 1,000 images. Random and SAR share the retrained checkpoint and four-step last-scale budget.
Autoregressive (AR) models based on next-scale prediction have emerged as a powerful tool for image generation, but they face a critical weakness: information inconsistencies between patches across timesteps introduced by progressive resolution scaling. These inconsistencies scatter guidance signals, causing them to drift away from salient regions within the image and leaving behind ambiguous, unfaithful features during sampling. We tackle this challenge with Information-Grounding Guidance (IGG), a novel framework that anchors guidance to semantically important tokens via an attention-based dynamic weighting formulation, consequently ensuring that guidance and semantic contents remain tightly aligned. Across both class-conditioned and text-to-image generation tasks, IGG delivers sharper, more coherent, and semantically grounded images, demonstrating its efficacy for correcting AR-based methods. Our code is available at https://github.com/dnngky/infoground-guidance.
Ky Dan Nguyen, Hoang Lam Tran, Anh-Dung Dinh +4
The University of Sydney, Sydney, New South Wales, Australia · The University of Western Australia, Perth, Western Australia, Australia
Autoregressive image generation has seen recent improvements with the introduction of chain-of-thought and reinforcement learning. However, current methods merely specify "What" details to depict by rewriting the input prompt, yet fundamentally fail to reason about "How" to structure the overall image. This inherent limitation gives rise to persistent issues, such as spatial ambiguity directly causing unrealistic object overlaps. To bridge this gap, we propose CoR-Painter, a novel framework that pioneers a "How-to-What" paradigm by introducing Constrained Reasoning to guide the autoregressive generation. Specifically, it first deduces "How to draw" by deriving a set of visual constraints from the input prompt, which explicitly govern spatial relationships, key attributes, and compositional rules. These constraints steer the subsequent generation of a detailed description "What to draw", providing a structurally sound and coherent basis for accurate visual synthesis. Additionally, we introduce a Dual-Objective GRPO strategy that specifically optimizes the textual constrained reasoning and visual projection processes to ensure the coherence and quality of the entire generation pipeline. Extensive experiments on T2I-CompBench, GenEval, and WISE demonstrate that our method achieves state-of-the-art performance, with significant improvements in spatial metrics (e.g., +5.41% on T2I-CompBench).
Ruxue Yan, Xubo Liu, Wenya Guo +3
College of Computer Science, Nankai University · Baidu Inc.
Autoregressive (AR) models have demonstrated strong potential in visual generation, offering superior performance with simple architectures and optimization objectives. However, existing methods are typically limited to single-modality conditions, e.g., text, restricting their applicability in real-world scenarios that demand image synthesis from diverse controls. In this work, we present OmniGen-AR, a unified autoregressive framework for Any-to-Image generation. By discretizing various visual conditions through a shared visual tokenizer and text prompts with a text tokenizer, OmniGen-AR supports a broad spectrum of conditional inputs within a single model, including text (text-to-image generation), spatial signals (segmentation-to-image and depth-to-image), and visual context (image editing, frame prediction, and text-to-video generation). To mitigate the risk of information leakage from condition tokens to content tokens, we introduce Disentangled Causal Attention (DCA), which separates the full-sequence causal mask into condition causal attention and content causal attention. It serves as a training-time regularizer without affecting the standard next-token prediction during inference. With this design, OmniGen-AR achieves new state-of-the-art or at least competitive results across a range of benchmark, e.g., 0.63 on GenEval and 80.02 on VBench, demonstrating its effectiveness in flexible and high-fidelity visual generation.
Junke Wang, Xun Wang, Qiushan Guo +4
Institute of Trustworthy Embodied AI, Fudan University · 2Shanghai Collaborative Innovation Center of Intelligent Visual Computing · 3Bytedance Seed +1