AlignVAR: Towards Globally Consistent Visual Autoregression for Image Super-Resolution

Cencen Liu; Dongyang Zhang; Wen Yin; Jielei Wang; Tianyu Li; Ji Guo; Wenbo Jiang; Guoqing Wang

cs.CVFeb 28, 2026

AlignVAR: Towards Globally Consistent Visual Autoregression for Image Super-Resolution

Authors: Cencen Liu, Dongyang Zhang, Wen Yin, Jielei Wang, Tianyu Li, Ji Guo, Wenbo Jiang, Guoqing Wang, +1 more

Organizations: University of Electronic Science and Technology of China · Ubiquitous Intelligence and Trusted Services Key Laboratory of Sichuan Province

Abstract

Visual autoregressive (VAR) models have recently emerged as a promising alternative for image generation, offering stable training, non-iterative inference, and high-fidelity synthesis through next-scale prediction. This encourages the exploration of VAR for image super-resolution (ISR), yet its application remains underexplored and faces two critical challenges: locality-biased attention, which fragments spatial structures, and residual-only supervision, which accumulates errors across scales, severely compromises global consistency of reconstructed images. To address these issues, we propose AlignVAR, a globally consistent visual autoregressive framework tailored for ISR, featuring two key components: (1) Spatial Consistency Autoregression (SCA), which applies an adaptive mask to reweight attention toward structurally correlated regions, thereby mitigating excessive locality and enhancing long-range dependencies; and (2) Hierarchical Consistency Constraint (HCC), which augments residual learning with full reconstruction supervision at each scale, exposing accumulated deviations early and stabilizing the coarse-to-fine refinement process. Extensive experiments demonstrate that AlignVAR consistently enhances structural coherence and perceptual fidelity over existing generative methods, while delivering over 10x faster inference with nearly 50% fewer parameters than leading diffusion-based approaches, establishing a new paradigm for efficient ISR.

Figures & tables

Figure 1: Comparison between the VARSR and AlignVAR. AlignVAR enhances VAR by introducing an adaptive consistency mask for intra-scale modeling and full reconstruction supervision for inter-scale alignment.

Figure 2: Comparison of attention distribution. Visualization of attention maps for VARSR and AlignVAR shows that VARSR exhibits highly localized attention concentrated in nearby regions, whereas AlignVAR captures broader contextual dependencies through the proposed Spatial Consistency Autoregression (SCA), thereby enhancing spatial coherence within each scale.

Figure 3: Spatial inconsistency results in texture discontinuities, structural distortions.

Figure 4: Hierarchical inconsistency results in color shifts and structural misalignment.

Figure 5: Overall architecture of the proposed AlignVAR . AlignVAR comprises two complementary components: a Spatial Consistency Autoregression (SCA) that performs scale-wise prediction and reweights intra-scale features using adaptive masks, and a Hierarchical Consistency Constraint (HCC) that jointly supervises residual and full representations to recalibrate inter-scale dependencies.

Figure 6: Comparison of multi-scale reconstructions. AlignVAR alleviates cumulative error propagation across scales and preserves more consistent structural and textural details than VARSR [ 28 ] .

Datasets	Metrics	GAN-based			Diffusion-based					VAR-based
Datasets	Metrics	BSRGAN	Real-ESR	SwinIR	LDM	StableSR	DiffBIR	PASD	UPSR	VARSR	AlignVAR
DIV2K-Val	PSNR $\uparrow$	24.42	24.30	23.77	21.66	23.26	23.49	23.59	24.87	24.41	24.35
	SSIM $\uparrow$	0.6164	0.6324	0.6186	0.4752	0.5670	0.5568	0.5899	0.6294	0.6189	0.6021
	LPIPS $\downarrow$	0.3511	0.3267	0.3910	0.4887	0.3228	0.3638	0.3611	0.3173	0.2985	0.2955
	DISTS $\downarrow$	0.2369	0.2245	0.2291	0.2693	0.2116	0.2177	0.2134	0.2132	0.2167	0.2162
	FID $\downarrow$	50.99	44.34	44.45	55.04	28.32	34.55	39.74	39.48	28.64	25.71

Table 1: Comparison with state-of-the-art methods on synthetic and real-world benchmarks. The best and second-best results are highlighted in bold red and underline blue , respectively.

Figure 7: Qualitative comparisons with representative methods.

Models	Params	Steps	Inference Time
StableSR	1409.1M	200	15.32s
DiffBIR	1900.4M	20	5.03s
PASD	1716.7M	50	5.94s
UPSR	1530.9M	5	2.79s
VARSR	1102.9M	10	0.52s
AlignVAR	1056.5M	10	0.43s

Table 2: Complexity comparison.

Metrics	RealSR			DrealSR
Metrics	w/o SCA	RI	SG	w/o SCA	RI	SG
PSNR $\uparrow$	26.77	26.67	26.11	29.32	28.68	28.54
LPIPS $\downarrow$	0.2847	0.2855	0.2871	0.2975	0.3077	0.3184
DISTS $\downarrow$	0.2079	0.2108	0.2123	0.2253	0.2289	0.2340
MANIQA $\uparrow$	0.4351	0.4435	0.4553	0.4285	0.4312	0.4685
MUSIQ $\uparrow$	66.74	67.21	68.53	62.75	63.06	63.83

Table 3: Ablation study of the SCA. “w/o” denotes removing SCA, “RI” indicates random input, and “SG” represents the structural guidance used in our model.

Figure 8: Comparison of supervision domains for HCC.

Metrics	RealSR		DrealSR
Metrics	w/o HCC	HCC	w/o HCC	HCC
PSNR $\uparrow$	25.85	26.11	28.23	28.54
LPIPS $\downarrow$	0.2898	0.2871	0.3208	0.3184
DISTS $\downarrow$	0.2175	0.2123	0.2411	0.2340
MANIQA $\uparrow$	0.4431	0.4553	0.4442	0.4685
MUSIQ $\uparrow$	67.06	68.53	62.85	63.83

Table 4: Ablation on the Hierarchical Consistency Constraint.

Settings	PSNR $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	MANIQA $\uparrow$	MUSIQ $\uparrow$
$\lambda=0.5$	25.83	0.2874	0.2184	0.4403	67.56
$\lambda=1.0$	26.11	0.2871	0.2123	0.4553	68.53
$\lambda=1.5$	26.21	0.2785	0.2107	0.4465	67.85
$\lambda=2.0$	26.35	0.2703	0.2079	0.4337	67.03

Table 5: Effect of the balancing coefficient

\lambda

on RealSR.

Figure 9: Analysis of the adaptive mask at different training stages. SG denotes the structural guidance extracted from the LR input.

Figure 10: Quantitative analysis of the Hierarchical Consistency Constraint on RealSR. Shown are the mean prediction errors across scales (left) and the MUSIQ scores under identical random perturbations injected at different scales (right).

Figure 11: Computational comparison among diffusion-based SR methods (DiffBIR [ 23 ] , LDM [ 29 ] , PASD [ 46 ] , UPSR [ 52 ] ), the VAR-based baseline (VARSR [ 28 ] ), and our AlignVAR. AlignVAR achieves high performance with substantially lower complexity than other models.

Datasets	Metrics	GAN-based			Diffusion-based					VAR-based
Datasets	Metrics	BSRGAN	Real-ESR	SwinIR	LDM	StableSR	DiffBIR	PASD	UPSR	VARSR	AlignVAR
RealLR200	NIQE $\downarrow$	4.3817	4.2048	4.2157	4.2533	4.2516	4.1715	4.9330	4.7606	4.4579	4.0617
	MANIQA $\uparrow$	0.5462	0.5582	0.3741	0.3049	0.5841	0.6066	0.5902	0.4206	0.4536	0.6237
	CLIPIQA $\uparrow$	0.5679	0.5389	0.5596	0.5253	0.6068	0.6797	0.6509	0.6397	0.6144	0.6734
	MUSIQ $\uparrow$	64.87	62.94	63.55	55.19	63.30	68.20	62.06	66.46	62.12	69.36

Table 6: Comparison with state-of-the-art methods on RealLR200. The best and second-best results are highlighted in bold red and underline blue , respectively.

Figure 12: User study results comparing the super-resolution outputs of different models.

Figure 13: Attention maps at different scales. AlignVAR expands the attention range, capturing dependencies over large areas.

Figure 14: Ablation results of SCA and HCC evaluated using Edge IoU [ 47 ] across different scales. The full AlignVAR achieves the highest edge alignment quality at all scales. Removing SCA results in the largest degradation, demonstrating its key role in maintaining spatial consistency and recovering fine-grained edges. HCC provides complementary improvements by enhancing hierarchical consistency, but its influence is comparatively weaker.

Figure 15: Impact of Spatial Consistency Autoregression (SCA) on attention maps. The attention range is significantly broader when SCA is applied.

Figure 16: Effect of the hyperparameter $\lambda$ on training dynamics. We compare four values of $\lambda$ (0.5, 1.0, 1.5, 2.0) in terms of loss and token prediction accuracy across different scales. A smaller $\lambda$ leads to faster convergence but lower accuracy, particularly at the middle scales (e.g., $6\times 6$ and $9\times 9$ ), while $\lambda=1.0$ achieves the best overall accuracy with stable convergence. Larger $\lambda$ values slow down loss convergence and degrade accuracy, indicating that excessively large consistency constraints hinder optimal learning.

Figure 17: Comparison with state-of-the-art methods on the RealLR200 dataset. The large image on the left is the output of AlignVAR, while the small patches on the right show enlarged crops from the LR input and competing models.

Figure 18: Additional qualitative comparisons. The large image on the left shows the output of AlignVAR, while the small patches on the right present enlarged crops from the LR input and competing methods.

Figure 19: Additional qualitative comparisons. The large image on the left shows the output of AlignVAR, while the small patches on the right present enlarged crops from the LR input and competing methods.

Explore similar work

VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution
Apr 23, 2026Yixuan Zhu, Shilin Ma, Haolin Wang +6Autoregressive Image Generation Knowledge Distillation
Tsinghua University
Adversarial Error Correction for Visual Autoregressive Generation
May 24, 2026Ligong Bi, Tao Huang, Jianyuan Guo +1Visual Autoregressive Models Autoregressive Image Generation
1Shanghai Jiao Tong University · 2City University of Hong Kong · 3The University of Sydney
Hierarchical Image Tokenization for Multi-Scale Image Super Resolution
May 14, 2026Isma Hadji, Enrique Sanchez, Adrian Bulat +2Image Tokenization Visual Autoregressive Models
Samsung AI Center Cambridge, UK · Technical University of Iasi, Romania · Queen Mary University of London, UK

Apr 23, 2026J/K move · Enter open · S save

VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution

Yixuan Zhu, Shilin Ma, Haolin Wang, Ao Li+5

Tsinghua University

Autoregressive Image Generation Knowledge Distillation Image Super-Resolution

Open paper AskPDF arXiv

Recent advancements in visual autoregressive models (VAR) have demonstrated their effectiveness in image generation, highlighting their potential for real-world image super-resolution (Real-ISR). However, adapting VAR for ISR presents critical challenges. The next-scale prediction mechanism, constrained by causal attention, fails to fully exploit global low-quality (LQ) context, resulting in blurry and inconsistent high-quality (HQ) outputs. Additionally, error accumulation in the iterative prediction severely degrades coherence in ISR task. To address these issues, we propose VARestorer, a simple yet effective distillation framework that transforms a pre-trained text-to-image VAR model into a one-step ISR model. By leveraging distribution matching, our method eliminates the need for iterative refinement, significantly reducing error propagation and inference time. Furthermore, we introduce pyramid image conditioning with cross-scale attention, which enables bidirectional scale-wise interactions and fully utilizes the input image information while adapting to the autoregressive mechanism. This prevents later LQ tokens from being overlooked in the transformer. By fine-tuning only 1.2% of the model parameters through parameter-efficient adapters, our method maintains the expressive power of the original VAR model while significantly enhancing efficiency. Extensive experiments show that VARestorer achieves state-of-the-art performance with 72.32 MUSIQ and 0.7669 CLIPIQA on DIV2K dataset, while accelerating inference by 10 times compared to conventional VAR inference.

Abstract

Explore similar work

VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution

Adversarial Error Correction for Visual Autoregressive Generation

Hierarchical Image Tokenization for Multi-Scale Image Super Resolution