Paper deep dive
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, Guoming Lu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 7/20/2026, 5:10:17 AM
Summary
The paper introduces AlignVAR, a globally consistent visual autoregressive framework for image super-resolution (ISR). It addresses two main challenges in existing VAR models: locality-biased attention causing spatial fragmentation and residual-only supervision leading to error accumulation across scales. AlignVAR proposes Spatial Consistency Autoregression (SCA) to enhance long-range dependencies via adaptive masking and Hierarchical Consistency Constraint (HCC) to stabilize coarse-to-fine refinement through full reconstruction supervision. Experiments show AlignVAR achieves superior structural coherence and perceptual fidelity with significantly faster inference and fewer parameters compared to diffusion-based methods.
Entities (10)
Relation Signals (8)
AlignVAR → uses → Hierarchical Consistency Constraint
confidence 95% · AlignVAR... featuring two key components: (2) Hierarchical Consistency Constraint (HCC)
AlignVAR → uses → Spatial Consistency Autoregression
confidence 95% · AlignVAR... featuring two key components: (1) Spatial Consistency Autoregression (SCA)
Hierarchical Consistency Constraint → addresses → residual-only supervision
confidence 90% · HCC... augments residual learning with full reconstruction supervision... exposing accumulated deviations early
Spatial Consistency Autoregression → mitigates → locality-biased attention
confidence 90% · SCA... mitigating excessive locality and enhancing long-range dependencies
AlignVAR → outperforms → VARSR
confidence 88% · AlignVAR consistently enhances structural coherence... over existing generative methods
AlignVAR → evaluatedon → DIV2K
confidence 85% · The synthetic validation set (DIV2K-Val) is built by randomly cropping 3K patches from DIV2K
AlignVAR → outperforms → Diffusion Models
confidence 85% · delivering over 10x faster inference with nearly 50% fewer parameters than leading diffusion-based approaches
AlignVAR → trainedon → LSDIR
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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.
Tags
Links
- Source: https://arxiv.org/abs/2603.00589v2
- Canonical: https://arxiv.org/abs/2603.00589v2
Trouble viewing inline? Open PDF directly →
Full Text
69,704 characters extracted from source content.
Expand or collapse full text
AlignVAR: Towards Globally Consistent Visual Autoregression for Image Super-Resolution Cencen Liu 1 ,Dongyang Zhang 1,2 ,Wen Yin 1 ,Jielei Wang 1,2 ,Tianyu Li 1 , Ji Guo 1 ,Wenbo Jiang 1 ,Guoqing Wang 1 ,Guoming Lu 1,2,∗ 1 University of Electronic Science and Technology of China 2 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 spa- tial structures, and residual-only supervision, which accu- mulates errors across scales, severely compromises global consistency of reconstructed images. To address these is- sues, 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 atten- tion toward structurally correlated regions, thereby mitigat- ing excessive locality and enhancing long-range dependen- cies; and (2) Hierarchical Consistency Constraint (HCC), which augments residual learning with full reconstruction supervision at each scale, exposing accumulated devia- tions early and stabilizing the coarse-to-fine refinement pro- cess. Extensive experiments demonstrate that AlignVAR consistently enhances structural coherence and perceptual fidelity over existing generative methods, while delivering over 10× faster inference with nearly 50% fewer parame- ters than leading diffusion-based approaches, establishing a new paradigm for efficient ISR. 1. Introduction Generative models, empowered by strong learned priors, have revolutionized the field of image super-resolution (ISR). Among them, Generative Adversarial Networks (GANs) [17, 19, 25] and Diffusion Models [23, 35, 46] have emerged as the two dominant paradigms. However, each of these approaches suffers from inherent limitations. GAN- based methods [17, 38, 50, 54], while enhancing perceptual EncoderNext-scale PredictionDecoder EncoderNext-scale PredictionDecoder Constrain (a) VARSR (b) Ours ......... ......... Figure 1. Comparison between the VARSR and AlignVAR. Align- VAR enhances VAR by introducing an adaptive consistency mask for intra-scale modeling and full reconstruction supervision for inter-scale alignment. realism, often exhibit training instability and tend to intro- duce visually inconsistent artifacts [44]. Diffusion-based approaches [7, 23, 35, 44, 46] exploit powerful generative priors to achieve high-fidelity reconstruction, yet their iter- ative denoising process incurs heavy computational costs, substantially limiting their practicality [4]. These challenges have motivated the exploration of alter- native generative paradigms for ISR, among which visual autoregressive (VAR) modeling has recently emerged as a promising direction [31, 42]. The coarse-to-fine prediction strategy of VAR naturally aligns with the hierarchical nature of ISR, which progressively restores image details across scales. The recent VARSR [28], as illustrated in Fig. 1, pio- neers the application of VAR to ISR, demonstrating its ini- tial feasibility. However, this seminal attempt also reveals a fundamental conflict: the model’s inherent locality bias and arXiv:2603.00589v2 [cs.CV] 5 Mar 2026 Figure 2. Comparison of attention distribution. Visualization of attention maps for VARSR and AlignVAR shows that VARSR exhibits highly localized attention concentrated in nearby re- gions, whereas AlignVAR captures broader contextual dependen- cies through the proposed Spatial Consistency Autoregression (SCA), thereby enhancing spatial coherence within each scale. the cumulative error propagation across scales jointly un- dermine the global consistency required for ISR, resulting in degraded global coherence in the reconstructed images. Building on this observation, we conduct a systematic analysis of the VARSR framework and identify two coupled root causes underlying its failure to maintain global consis- tency. The first is Spatial Inconsistency. The self-attention mechanism in VAR models exhibits a strong locality bias, with attention weights concentrated almost exclusively on adjacent regions, as illustrated in Fig. 2. This restricted re- ceptive field limits the integration of global context, result- ing in spatially disjoint artifacts such as fragmented textures and structural distortions, as shown in Fig. 3. The second is Hierarchical Inconsistency, which we analyze by inject- ing random perturbations into different scales and observ- ing their influence on the final reconstruction, as shown in Fig. 4. These perturbations induce color shifts and struc- tural misalignments, indicating that residual-only supervi- sion allows small prediction errors from coarser scales to propagate and amplify through the hierarchy [15]. We argue that these issues stem from a common origin: the lack of ex- plicit consistency constraints in both the intra-scale (spatial) and inter-scale (hierarchical) dimensions. To address these limitations, we propose AlignVAR, a visual autoregressive framework designed to achieve glob- ally consistent ISR. As shown in Fig. 1, AlignVAR intro- duces two complementary mechanisms that collaboratively promote coherence within and across scales by reweighting spatial attention and recalibrating hierarchical dependen- cies. Specifically, the Spatial Consistency Autoregression (SCA) adaptively reweights attention to emphasize struc- turally correlated regions rather than local neighborhoods, enabling the model to aggregate long-range context and maintain spatial continuity. Meanwhile, the Hierarchical Consistency Constraint (HCC) recalibrates cross-scale de- pendencies through a hierarchical supervision objective that enforces both residual and full-scale consistency between predictions and ground-truth representations. By allowing w/o HCCLatent-HCCPixel-HCCHR LR Scale 8Scale 10Scale 6Scale 8Scale 10Scale 6 VARSR Ours VARSR Ours VARSR Ours HRHR Figure 3. Spatial inconsistency results in texture discontinuities, structural distortions. Figure 4. Hierarchical inconsistency results in color shifts and structural misalignment. each scale to correct contextual deviations before they prop- agate, HCC effectively suppresses error accumulation and stabilizes refinement throughout the hierarchy. SCA and HCC form a unified consistency-driven autoregressive for- mulation that restores fine textures while preserving large- scale structural fidelity, yielding reconstructions that are both perceptually accurate and globally coherent. In summary, we make the following key contributions: • We revisit visual autoregression for ISR and identify two fundamental sources of inconsistency in existing VAR- based frameworks: spatial inconsistency arising from lo- cality bias and hierarchical inconsistency caused by cu- mulative error propagation. • We propose AlignVAR, a globally consistent VAR frame- work that enhances spatial coherence and hierarchical alignment through two complementary components: Spa- tial Consistency Autoregression (SCA) and Hierarchical Consistency Constraint (HCC). • Extensive experiments demonstrate that AlignVAR sub- stantially improves reconstruction coherence and percep- tual quality, establishing a strong benchmark for autore- gressive real-world ISR. 2. Related Works Image super-resolution. Early image super-resolution methods learn deterministic mappings from low-resolution to high-resolution images under simplified degradations such as bicubic downsampling, limiting their generalization to real-world scenarios [8, 10, 21, 22, 24, 55]. To overcome this, GAN-based approaches employ adversarial losses to approximate natural image distributions [16, 17, 38, 50, 54]. Methods like BSRGAN [51] and Real-ESRGAN [39] in- troduce blind degradation modeling for robust real-world recovery, but often suffer from unstable training and unnat- ural artifacts [44]. Recently, diffusion-based SR methods such as DiffBIR [23], StableSR [35], SeeSR [44], PASD [46], and PiSA-SR [30] leverage generative priors from pre- trained diffusion models to enhance perceptual fidelity. De- spite their strong visual performance, these models rely on iterative denoising, which incurs high computational cost and may produce hallucinated or inconsistent details [4, 27]. Visual autoregressive models. Autoregression, moti- vated by its success in large language models [1, 32], has recently gained traction in visual generation tasks. Most approaches discretize latent features using vector quantiz- ers [18, 33] and predict tokens sequentially [6, 11, 40, 42, 48, 49]. However, next-token prediction over flattened sequences often breaks spatial structure, which makes it difficult to generate coherent high-resolution content [31]. Visual autoregressive modeling [31] addresses this limita- tion through next-scale prediction, reconstructing images progressively across multiple scales and achieving strong generative performance. This paradigm has also been ex- plored in image restoration [37] and further extended to ISR through VARSR [28], which reconstructs HR images in a coarse-to-fine manner. Nevertheless, VARSR still suffers from locality bias and cumulative error propagation, lead- ing to both spatial and hierarchical inconsistency. 3. Preliminaries In next-scale visual autoregression, image reconstruction is performed hierarchically within the latent space of a vector- quantized variational autoencoder (VQ-VAE) [33]. At each scale k, the latent feature f k is computed as the residual be- tween the target latent f and the upsampled reconstructions from all previous coarser scales: f k = f − k−1 X m=1 upsample lookup(V,r m ) ,(1) whereV denotes the shared codebook, and lookup(V,r m ) retrieves the embedded vectors corresponding to tokens r m . Each latent position is quantized to the nearest codebook entry, yielding the discrete token map r k : r k (i,j) = arg min v∈[|V|] lookup(V,v)− f k (i,j) 2 2 .(2) Across scales, the overall autoregressive process follows a coarse-to-fine factorization: p(r 1 ,r 2 ,...,r K ) = K Y k=1 p θ (r k | r 1 ,r 2 ,...,r k−1 ,c), (3) where c represents the conditional latent derived from the low-resolution input. 4. Methods 4.1. Motivation and Empirical Observation We analyze the spatial behavior of visual autoregression (VAR) in image super-resolution by visualizing its atten- tion distribution. As shown in Fig. 2, VARSR [28] exhibits highly localized attention responses, where most attention mass is confined to narrow regions, revealing a strong local bias. This bias limits long-range contextual aggregation and weakens spatial correlations among distant yet related struc- tures. These observations indicate that sequential tokeniza- tion and limited receptive field lead to a non-conservative spatial information flow, resulting in spatial inconsistency within each scale. This motivates our first question: Q1: How can we mitigate the local bias of VAR to achieve spatial consistency within each scale? Hierarchical inconsistency caused by error accumula- tion. Beyond spatial coherence, the hierarchical predic- tion process of VAR inherently suffers from cumulative pre- diction errors, as shown in Fig. 4. In the next-scale predic- tion paradigm, each latent r k is estimated based on imper- fect outputs from preceding scales r 1:k−1 . Any deviation in coarse-scale predictions shifts the conditional distribution p(r k | r 1:k−1 ), leading to a compounding effect in which errors are propagated and amplified across scales. We at- tribute this problem to residual-only supervision used in ex- isting VAR frameworks, which leaves intermediate latent representations under-constrained and destabilizes the gen- erative hierarchy. This motivates our second question: Q2: How can we recalibrate intermediate predictions to maintain hierarchical consistency across scales? 4.2. Overview Motivated by the two issues identified above, we pro- pose AlignVAR, a globally consistent visual autoregressive framework for ISR. The overall architecture of AlignVAR is illustrated in Fig. 5. Briefly, AlignVAR is composed of two primary modules: Spatial Consistency Autoregression (SCA) and Hierarchical Consistency Constraint (HCC). The SCA expands the effective receptive field and applies a learnable structure-aware mask to mitigate local attention bias (§4.3), while the HCC enforces full-scale latent align- ment to recalibrate intermediate predictions and suppress cumulative drift across scales (§4.4). Together, these de- signs enable AlignVAR to produce perceptually faithful and structurally coherent high-resolution reconstructions. 4.3. Spatial Consistency Autoregression (SCA) The core idea of SCA is to enhance spatial coherence within the autoregressive process. Rather than relying on order-based attention with an inherently local bias, SCA introduces a structure-aware conditioning principle that reweights contextual dependencies. ... ... Scalable Visual Autoregressive Transformer Cond Encoder Teacher-forcing Input HR LR ...... ... ... Condition Degradation Pipeline Hierarchical Consistency Constraint VA E Encoder ❄ Latent from Quantizated Full HR ... Frozen Module Trainable Module Matrix Addition Latent Accumulation of Predicted Residuals ❄ ...... ... ... Laplace Filter Down sample ... ... ... Multi-scale Adaptive Masks Structural Guidance Mask Generator ... ...... Adaptive Mask Tokens Mask Modulation Multi-scale Quantization Train only Inference only HCC LossLatent of Ground-truth Upsample Upsample Spatial Consistency Autoregression VA E Decoder ❄ SR Output G G 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. Specifically, SCA preserves the coarse-to-fine au- toregressive structure but replaces the plain context (r 1 ,...,r k−1 ) with a structure-aware counterpart. Let ̃r k denote the reweighted tokens at scale k. The joint distribu- tion under SCA becomes p SCA ( ̃r 1 ,..., ̃r K ) = K Y k=1 p θ ̃r k | ̃r 1 , ̃r 2 ,..., ̃r k−1 ,c , (4) where the next-scale prediction at level k is conditioned on structure-aligned representations of all coarser scales. SCA thus remains defined only across scales, while spatial con- sistency of the context is enforced through structure-aware reweighted tokens ̃r 1:k−1 . Structure-aware reweighting field. As shown in the bot- tom middle part of Fig. 5, we use structure-aware guid- ance extracted from the low-resolution input I LR to capture structural cues corresponding to potential edge or texture regions: s = Laplacian(I LR ) , ̄s k = norm Down k (s) , (5) where Laplacian represents a Laplacian operator [5], cho- sen for its sensitivity to second-order structural changes, enabling the guidance map to highlight edges for effective spatial reweighting. Down k (·) denotes scale-specific down- sampling that matches the spatial resolution of the k-th la- tent map, and norm(·) rescales values to [0, 1]. By normalizing across scales, the structural guidance ̄s k remains comparable in magnitude and semantics, allow- ing the model to leverage multi-scale geometric consistency rather than relying solely on local correlations. Subsequently, we use a lightweight MLP-based mask generatorM φ to predict a spatial modulation field m k from the autoregressive tokens r k and the corresponding struc- tural guidance ̄s k : m k = σ M φ ([r k , ̄s k ]) ,(6) where [· ] denotes channel-wise concatenation and σ is a sigmoid function. This modulation field m k serves as a structure-aware reweighting map that adaptively adjusts the token responses according to the reliability of the under- lying structure. Regions with clear geometric cues are as- signed higher weights, encouraging the model to attend to stable, well-defined features, while uncertain or textureless areas are softly suppressed to reduce local noise amplifi- cation. Finally, the structure-aware reweighted tokens are produced through a learnable spatial field via token gating: ̃r k = (1 + m k )⊙ r k ,(7) with ⊙ denoting element-wise multiplication. This oper- ation can be viewed as a spatially adaptive gain control that modulates the strength of autoregressive context prop- agation. By reinforcing structure-aligned activations and suppressing unstable ones, the reweighting field guides the model to preferentially propagate information along reliable structural paths. This selectively amplifies long-range cor- relations between semantically related regions, effectively mitigating local bias and enabling broader contextual ag- gregation within each scale. While SCA enforces spatial coherence within each scale, the hierarchical dependencies across scales remain unregu- larized. Section 4.4 introduces the Hierarchical Consistency Constraint to align inter-scale representations and suppress error propagation through the hierarchy. w/o HCCLatent-HCCPixel-HCCHR LR Scale 8Scale 10Scale 6Scale 8Scale 10Scale 6 VARSR Ours 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]. 4.4. Hierarchical Consistency Constraint (HCC) In the original next-scale autoregressive formulation, the Cross Entropy (CE) loss [31] supervises only the resid- ual tokens predicted at each scale. While effective for lo- cal refinement, this residual-only supervision cannot cor- rect errors accumulated from coarser levels, allowing scale- wise biases to propagate upward. To address this issue, we introduce HCC, which provides full-scale latent supervi- sion to explicitly recalibrate hierarchical dependencies. By aligning predicted latent representations with their complete multi-scale ground truths, HCC enforces global consistency across the entire autoregressive hierarchy. Full-scale representation. Given a high-resolution image I HR ∈R C×H K ×W K , the VAE encoderE produces its latent feature representation z =E (I HR )∈R C×H K ×W K .(8) To construct scale-specific ground truths for hierarchical su- pervision, we spatially downsample z to each target resolu- tion S k and quantize the result into discrete tokens: u k gt =Q Down(z,S k ) , S k ∈S 1 ,S 2 ,...,S K , (9) where Down(·) denotes spatial downsampling to the reso- lution of scale k, andQ(·) denotes the shared vector quan- tizer. Unlike the residual tokens used in the standard CE loss, u k gt represents the complete latent state at each scale, thus providing a stronger supervision target. HCC loss. During autoregressive reconstruction, the model predicts residual tokens ˆr k pred for each scale. The cor- responding cumulative (full-scale) latent prediction is ob- tained by aggregating residuals from all previous scales: ˆu k pred = ˆu k−1 pred + ˆr k pred ,ˆu (0) pred =0.(10) We design the HCC loss to impose a multi-scale hierarchi- cal alignment constraint between these cumulative predic- tions and the corresponding ground-truth latents: L HCC = K X k=1 ˆu k pred − u k gt 2 2 .(11) This inter-scale calibration prevents error propagation and aligns the semantic content across the hierarchy. By extending the supervision from residuals to full la- tent representations, the HCC loss bridges the semantic gap between hierarchical scales and stabilizes the coarse-to-fine refinement process. As shown in Fig. 6, AlignVAR recon- structs finer and more consistent textures at higher scales compared to VARSR [28], indicating that the proposed re- calibration effectively alleviates the accumulation and prop- agation of local errors. Together with SCA’s spatial regu- larization, it establishes a unified mechanism that preserves intra-scale coherence and inter-scale consistency through- out the autoregressive super-resolution process. 4.5. Training Objective Our training process is performed under the teacher-forcing paradigm [31], where the prediction at each scale is con- ditioned on the reweighted ground-truth tokens ̃r 1:k−1 gt and the low-resolution latent feature c, providing a stable and reliable supervision path throughout the entire multi-scale prediction procedure. The autoregressive predictor models a categorical distribution over the codebook entries at each spatial position (i,j), optimized via a standard CE objective commonly used in modern generative modeling: L CE =− K X k=1 X (i,j) logp θ r k gt (i,j)| ̃r 1:k−1 gt ,c .(12) Finally, the overall objective of AlignVAR combines both intra-scale and inter-scale learning signals, ensuring that each component contributes meaningfully to the reconstruc- tion quality: L total =L CE + λL HCC ,(13) where λ is a hyperparameter that balances the two compo- nents. Joint optimization of M φ and p θ encourages spa- tially coherent predictions at each scale and hierarchically consistent dependencies across the reconstruction process, ultimately improving both stability and performance. 5. Experimental Results 5.1. Experimental Setups Datasets. AlignVAR is trained on LSDIR [20] and the first 10K face images from FFHQ [13], where low- resolution counterparts are synthesized using the Real- ESRGAN degradation pipeline [39]. For evaluation, we adopt both synthetic and real-world benchmarks. The syn- thetic validation set (DIV2K-Val) is built by randomly crop- ping 3K patches from DIV2K [3], while real-world evalua- tion is conducted on DRealSR [43] and RealSR [36]. Fol- lowing [31], all HR images are standardized to 512× 512 and their LR inputs to 128× 128. Table 1. Comparison with state-of-the-art methods on synthetic and real-world benchmarks. The best and second-best results are high- lighted in bold red and underline blue, respectively. DatasetsMetrics GAN-basedDiffusion-basedVAR-based BSRGANReal-ESRSwinIRLDMStableSRDiffBIRPASDUPSRVARSRAlignVAR DIV2K-Val PSNR↑24.4224.3023.7721.6623.2623.4923.5924.8724.4124.35 SSIM↑0.61640.63240.61860.47520.56700.55680.58990.62940.61890.6021 LPIPS↓0.35110.32670.39100.48870.32280.36380.36110.31730.29850.2955 DISTS↓0.23690.22450.22910.26930.21160.21770.21340.21320.21670.2162 FID↓50.9944.3444.4555.0428.3234.5539.7439.4828.6425.71 MANIQA↑0.35470.37560.34110.35890.41730.45980.44400.36350.41370.4665 CLIPIQA↑0.52530.52050.52130.55700.67520.67310.65730.57480.63120.6754 MUSIQ↑60.1859.7657.2157.4665.1965.5766.5862.3866.8867.32 RealSR PSNR↑26.3825.6825.8825.6624.6924.9425.2125.9726.0826.11 SSIM↑0.76510.76140.76710.69340.70900.66640.71400.74490.73810.7125 LPIPS↓0.26560.27100.26140.33670.30030.34850.29860.28620.27770.2871 DISTS↓0.21240.20600.20610.23240.21340.22570.21250.20830.21110.2123 FID↓141.25135.14132.80133.34131.72127.59139.42140.11118.84138.81 MANIQA↑0.37630.37360.35610.33750.41670.43780.44180.38930.43160.4553 CLIPIQA↑0.51140.44870.44330.60530.62000.63960.60090.58560.59530.6784 MUSIQ↑63.2860.3759.2856.3265.2564.3266.6164.7966.6568.53 DRealSR PSNR↑28.7028.6128.2027.7827.8726.5727.4529.2128.8428.54 SSIM↑0.80280.80520.79830.71520.74270.65160.75390.80180.77350.7682 LPIPS↓0.28580.28190.28300.37450.33330.45370.33310.28720.31020.3184 DISTS↓0.21440.20890.21030.24170.22970.27240.23220.20500.23110.2340 FID↓155.62147.66146.38164.87148.18160.67173.40157.83141.99145.44 MANIQA↑0.34410.34350.33110.33420.38970.46020.45510.33390.42250.4685 CLIPIQA↑0.50610.45250.45220.59840.63210.64450.63650.55370.63950.6534 MUSIQ↑57.1654.2753.0151.3758.7261.0663.6955.8862.6663.83 Implementation details. AlignVAR is trained using a scale-wise autoregressive transformer backbone consisting of 24 transformer blocks. The model is initialized from pre- trained VAR weights [31] for faster convergence. Training uses AdamW [2] with a batch size of 32, weight decay of 5×10 −2 , and an initial learning rate of 5×10 −5 decayed via cosine annealing. The model is optimized for 100 epochs, with the loss balancing coefficient λ in Eq. 13 set to 1.0. All experiments are conducted on 8 NVIDIA H100 GPUs. Evaluation metrics. To comprehensively evaluate both fidelity and perceptual quality, we adopt a diverse set of reference-based and no-reference metrics.For fidelity, PSNR and SSIM [41] are computed on the Y channel of the YCbCr color space. Perceptual similarity is measured using LPIPS [53] and DISTS [9], while distributional alignment is assessed with FID [12]. We further employ no-reference metrics, including MANIQA [45], MUSIQ [14], and CLIP- IQA [34], to evaluate visual naturalness. Compared models. We compare AlignVAR with state- of-the-art (SOTA) methods in two categories. The first includes GAN-based methods:BSRGAN [51], Real- ESRGAN [39], and SwinIR-GAN [21]. The second covers diffusion-based methods: LDM [29], StableSR [35], Diff- BIR [23], PASD [46], and UPSR [52]. We also include the recent autoregressive method VARSR [28], retrained on our dataset for fairness. Other competing models are evaluated using official implementations and pretrained weights. 5.2. Comparison with SOTA Quantitative comparisons. As shown in Table 1, Align- VAR performs strongly across both synthetic and real- world benchmarks, substantially improving overall percep- tual quality. On the DIV2K-Val dataset, it outperforms GAN-based and diffusion-based methods in all perceptual metrics, achieving the lowest FID of 25.71 and the best LPIPS of 0.2955 among all compared models. On Re- alSR, AlignVAR significantly increases MUSIQ from 66.65 to 68.53 and CLIPIQA from 0.5953 to 0.6784 compared to VARSR [28], demonstrating its consistent effectiveness. Although AlignVAR does not achieve the highest fidelity metrics, this is expected because fine details in LR im- ages are severely degraded and cannot be fully recovered by any existing method. AlignVAR consistently produces high-quality super-resolution results that better align with human visual perception, offering perceptually realistic, vi- sually coherent, and structurally consistent reconstructions. Qualitative comparisons. Fig. 7 presents visual compar- isons with representative ISR methods. Some artifacts high- light common shortcomings of existing paradigms: GAN- based models often cause local distortions and jagged edges, while diffusion-based models may hallucinate tex- tures and weaken structural alignment. In contrast, Align- VAR reconstructs sharp edges, coherent textures, and natu- ral color transitions, better aligning with human visual per- ception. Notably, when true high-frequency details are ir- LRBSRGAN Real-ESRGAN PASDUPSR VARSR AlignVARHR Figure 7. Qualitative comparisons with representative methods. Table 2. Complexity comparison. ModelsParamsStepsInference Time StableSR1409.1M20015.32s DiffBIR1900.4M205.03s PASD 1716.7M505.94s UPSR1530.9M52.79s VARSR 1102.9M100.52s AlignVAR1056.5M100.43s 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. Metrics RealSRDrealSR w/o SCARISGw/o SCARISG PSNR↑26.7726.6726.1129.3228.6828.54 LPIPS↓0.28470.28550.28710.29750.30770.3184 DISTS↓ 0.20790.21080.21230.22530.22890.2340 MANIQA↑0.43510.44350.45530.42850.43120.4685 MUSIQ↑ 66.7467.2168.5362.7563.0663.83 recoverably lost in the low-resolution input, AlignVAR pre- serves the recoverable structures and generates perceptually plausible textures. This explains its clearly superior no- reference metric scores, even if the reference metric scores in Table 1 are not correspondingly higher. Complexity comparisons. As shown in Table 2, Align- VAR achieves significantly higher efficiency than diffusion- based models and VARSR [28]. Diffusion-based methods require iterative refinement, leading to long inference times even with a limited number of sampling steps. In contrast, AlignVAR reconstructs a 512× 512 image in only 0.43 sec- onds, which is over 10× faster than PASD and more than 5× faster than the 5-step UPSR. This efficiency stems from the scale-wise prediction strategy, where early scales in- w/o HCCLatent-HCCPixel-HCCHR LR Scale 8Scale 10Scale 6Scale 8Scale 10Scale 6 VARSR Ours Figure 8. Comparison of supervision domains for HCC. volve fewer tokens and thus incur lower computational cost. Compared with the autoregressive baseline VARSR, which incorporates an additional diffusion refiner at the expense of increased parameters and latency, AlignVAR removes this component and adopts a lightweight mask generator that introduces negligible overhead. Thus, AlignVAR achieves better overall performance with fewer parameters and faster inference. Further analysis of the efficiency advantage of AlignVAR is provided in the Supplementary Material. 5.3. Ablation Study and Analysis Effectiveness of SCA. We evaluate the effectiveness of SCA on the RealSR and DRealSR datasets, as shown in Ta- ble 3. Three configurations are compared: removing SCA, using random input for the mask generator (RI), and apply- ing the proposed structural guidance (SG). Removing SCA slightly improves fidelity metrics but noticeably degrades perceptual scores, while random input leads to unstable tex- tures and weakened visual coherence. These results demon- strate that SCA effectively exploits structural guidance to enhance spatial consistency and achieves a better balance between fidelity and perceptual quality. Effectiveness of HCC. We evaluate the impact of HCC on the RealSR and DRealSR datasets, as shown in Ta- Table 4. Ablation on the Hierarchical Consistency Constraint. Metrics RealSRDrealSR w/o HCCHCCw/o HCCHCC PSNR↑25.8526.1128.2328.54 LPIPS↓0.28980.28710.32080.3184 DISTS↓ 0.21750.21230.24110.2340 MANIQA↑0.44310.45530.44420.4685 MUSIQ↑67.0668.5362.8563.83 Table 5. Effect of the balancing coefficient λ on RealSR. SettingsPSNR↑LPIPS↓DISTS↓MANIQA↑MUSIQ↑ λ = 0.525.830.28740.21840.440367.56 λ = 1.026.110.28710.21230.455368.53 λ = 1.5 26.210.27850.21070.446567.85 λ = 2.026.350.27030.20790.433767.03 ble 4. Introducing HCC consistently improves both fidelity and perceptual metrics, indicating that aligning full-scale latent representations strengthens cross-scale information flow and stabilizes hierarchical prediction. We further com- pare applying this constraint in latent space versus in pixel space and observe that latent space supervision achieves no- ticeably better perceptual coherence, as illustrated in Fig. 8. Overall, HCC effectively suppresses accumulated errors and enhances inter-scale consistency, yielding reconstruc- tions that are both more accurate and visually coherent. Influence of the balancing coefficient. We study the ef- fect of the balancing coefficient λ in Eq. 13 on the RealSR dataset, as shown in Table 5. When λ increases from 0.5 to 1.0, both perceptual and fidelity metrics improve, with the best perceptual quality achieved at λ = 1.0. Further in- creasing λ beyond 1.0 slightly enhances distortion-oriented scores but leads to a decrease in perceptual quality. These results suggest that λ = 1.0 provides better overall per- formance in terms of perceptual quality while maintaining competitive fidelity scores. Analysis of the Adaptive Mask. Fig. 9 visualizes the structural guidance, the corresponding initial and learned masks, and the reweighted feature map. The structural guid- ance captures coarse edges and contours, which are trans- ferred to the initial mask as a spatial bias. After training, the learned mask becomes sharper and more selective, concen- trating on semantic boundaries such as windmill blades and Chinese characters while suppressing activations in smooth background regions. The reweighted feature map highlights these structural areas with clearer and more coherent re- sponses, suggesting that the mask adaptively emphasizes perceptually important regions. Overall, this process en- ables more structure-aware feature interaction and improves spatial consistency in the reconstructed images. Analysis of HCC. To further understand the effect of HCC, we analyze its impact on hierarchical recalibration and robustness to perturbations. Fig. 10 (left) plots the mean LR InputInitial MaskFeature MapSG Learned Mask Early StageConverged Stage 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). squared error (MSE) at each prediction scale. AlignVAR without HCC already reduces overall errors compared with VARSR, and adding HCC further lowers the MSE at early and middle scales, indicating that HCC serves as a coarse- to-fine recalibration mechanism that corrects global struc- tures before finer refinement. For robustness evaluation, we inject identical random perturbations into latent features at scales 2, 5, and 8 and report MUSIQ scores in Fig. 10 (right). AlignVAR shows the smallest MUSIQ degradation under all noise levels, confirming that HCC improves inter- scale stability against cumulative perturbations. 6. Conclusion We present AlignVAR, a globally consistent visual autore- gressive framework for image super-resolution. Through a detailed and comprehensive technical analysis of incon- sistency issues in existing VAR models, we identify spatial local bias and hierarchical error propagation as the main bottlenecks severely hindering coherent and accurate re- construction. To alleviate these issues, we introduce Spa- tial Consistency Autoregression to adaptively reweight fine- grained intra-scale attention and Hierarchical Consistency Constraint to recalibrate multi-level inter-scale dependen- cies through joint residual and full-scale supervision. Ex- tensive experiments on both synthetic and real-world bench- marks consistently demonstrate that AlignVAR achieves an excellent overall balance between fidelity and perceptual quality. This work provides a new perspective for achiev- ing globally coherent and reliable visual autoregression for high-quality image super-resolution. References [1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 [2] Kingma DP Ba J Adam et al. A method for stochastic op- timization. arXiv preprint arXiv:1412.6980, 1412(6), 2014. 6 [3] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPR workshops, pages 126–135, 2017. 5, 2 [4] Sumukh K Aithal, Pratyush Maini, Zachary Lipton, and J Zico Kolter.Understanding hallucinations in diffusion models through mode interpolation. NeurIPS, 37:134614– 134644, 2024. 1, 3 [5] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671–679. Elsevier, 1987. 4 [6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, pages 11315–11325, 2022. 3 [7] Bin Chen, Gehui Li, Rongyuan Wu, Xindong Zhang, Jie Chen, Jian Zhang, and Lei Zhang. Adversarial diffusion compression for real-world image super-resolution. In Pro- ceedings of the Computer Vision and Pattern Recognition Conference, pages 28208–28220, 2025. 1 [8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super- resolution transformer. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 22367–22377, 2023. 2 [9] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and ma- chine intelligence, 44(5):2567–2581, 2020. 6 [10] Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Learning a deep convolutional network for image super-resolution. In European conference on computer vi- sion, pages 184–199. Springer, 2014. 2 [11] Baisong Guo, Xiaoyun Zhang, Haoning Wu, Yu Wang, Ya Zhang, and Yan-Feng Wang. Lar-sr: A local autoregres- sive model for image super-resolution. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1899–1908, 2022. 3 [12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. NeurIPS, 30, 2017. 6 [13] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4401–4410, 2019. 5 [14] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, pages 5128–5137, 2021. 6, 3 [15] Hermann Kumbong, Xian Liu, Tsung-Yi Lin, Ming-Yu Liu, Xihui Liu, Ziwei Liu, Daniel Y Fu, Christopher Re, and David W Romero. Hmar: Efficient hierarchical masked auto- regressive image generation. In CVPR, pages 2535–2544, 2025. 2 [16] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming- Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, pages 624–632, 2017. 2 [17] Christian Ledig, Lucas Theis, Ferenc Husz ́ ar, Jose Caballero, Andrew P. Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super- resolution using a generative adversarial network.2017 IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), pages 105–114, 2016. 1, 2 [18] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han.Autoregressive image generation using residual quantization. In CVPR, pages 11523–11532, 2022. 3 [19] Wenbo Li, Kun Zhou, Lu Qi, Liying Lu, and Jiangbo Lu. Best-buddy gans for highly detailed image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 1412–1420, 2022. 1 [20] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Deman- dolx, et al. Lsdir: A large scale dataset for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1775–1787, 2023. 5 [21] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration us- ing swin transformer. In ICCV, pages 1833–1844, 2021. 2, 6, 3 [22] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE confer- ence on computer vision and pattern recognition workshops, pages 136–144, 2017. 2 [23] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff- bir: Toward blind image restoration with generative diffusion prior. In ECCV, pages 430–448. Springer, 2024. 1, 3, 6, 2 [24] Cencen Liu, Dongyang Zhang, Guoming Lu, Wen Yin, Jielei Wang, and Guangchun Luo.Srmamba-t: Exploring the hybrid mamba-transformer network for single image super- resolution. Neurocomput., 624(C), 2025. 2 [25] Chenxi Ma. Uncertainty-aware gan for single image super resolution. In Proceedings of the AAAI Conference on Arti- ficial Intelligence, pages 4071–4079, 2024. 1 [26] Anish Mittal, Anush K. Moorthy, and Alan C. Bovik. No- reference image quality assessment in the spatial domain. IEEE Signal Processing Letters, 19(12):813–816, 2012. 3 [27] Supreeth Narasimhaswamy, Uttaran Bhattacharya, Xiang Chen, Ishita Dasgupta, Saayan Mitra, and Minh Hoai. Hand- iffuser: Text-to-image generation with realistic hand appear- ances. In CVPR, pages 2468–2479, 2024. 3 [28] Yunpeng Qu, Kun Yuan, Jinhua Hao, Kai Zhao, Qizhi Xie, Ming Sun, and Chao Zhou. Visual autoregressive modeling for image super-resolution. In ICML, 2025. 1, 3, 5, 6, 7, 2 [29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ̈ orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, pages 10684– 10695, 2022. 6, 2, 3 [30] Lingchen Sun, Rongyuan Wu, Zhiyuan Ma, Shuaizheng Liu, Qiaosi Yi, and Lei Zhang. Pixel-level and semantic-level ad- justable super-resolution: A dual-lora approach. In Proceed- ings of the Computer Vision and Pattern Recognition Con- ference, pages 2333–2343, 2025. 3 [31] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image gen- eration via next-scale prediction. NeurIPS, 37:84839–84865, 2024. 1, 3, 5, 6, 2 [32] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth ́ e Lacroix, Baptiste Rozi ` ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 [33] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3, 1 [34] Jie Wang, Yixin Li, Yifan Wang, Siyuan Chen, Qifan Wang, and Zhangyang Wang. Clip-iqa: No-reference image quality assessment with contrastive pre-training. In CVPR, pages 1502–1511, 2023. 6, 2, 3 [35] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12):5929–5949, 2024. 1, 3, 6, 2 [36] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12):5929–5949, 2024. 5, 2 [37] Siyang Wang and Feng Zhao. Varformer: Adapting var’s generative prior for image restoration. arXiv e-prints, pages arXiv–2412, 2024. 3 [38] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: En- hanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 1, 2 [39] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV, pages 1905–1914, 2021. 2, 5, 6, 3 [40] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3 [41] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 6 [42] Hongyang Wei, Shuaizheng Liu, Chun Yuan, and Lei Zhang. Perceive, understand and restore: Real-world image super- resolution with autoregressive multimodal generative mod- els. arXiv preprint arXiv:2503.11073, 2025. 1, 3 [43] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qix- iang Ye, Wangmeng Zuo, and Liang Lin.Component divide-and-conquer for real-world image super-resolution. In ECCV, pages 101–117. Springer, 2020. 5, 2 [44] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, pages 25456– 25467, 2024. 1, 3 [45] Sidi Yang, Tianhe Wu, Shuwei Shi, Han Gong, Jing Cao, Bin Yang, Xu Jia, and Xiaogang Zhu. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In CVPR, pages 1191–1200, 2022. 6, 3 [46] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic im- age super-resolution and personalized stylization. In ECCV, pages 74–91. Springer, 2024. 1, 3, 6, 2, 5 [47] Jiahui Yu, Yuning Jiang, Zhangyang Wang, Zhimin Cao, and Thomas Huang. Unitbox: An advanced object detection net- work. In ACM M, page 516–520, New York, NY, USA, 2016. Association for Computing Machinery. 3, 4 [48] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gun- jan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin- fei Yang, Burcu Karagol Ayan, et al. Scaling autoregres- sive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 3 [49] Kaiwen Yu, Qi He, Gang Wu, and Zhijin Qin.Partial sampling-based semantic communications. IEEE Transac- tions on Communications, 73(10):9059–9070, 2025. 3 [50] Yuan Yuan, Siyuan Liu, Jiawei Zhang, Yongbing Zhang, Chao Dong, and Liang Lin. Unsupervised image super- resolution using cycle-in-cycle generative adversarial net- works. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 701–710, 2018. 1, 2 [51] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timo- fte. Designing a practical degradation model for deep blind image super-resolution. In ICCV, pages 4791–4800, 2021. 2, 6, 3, 5 [52] Leheng Zhang, Weiyi You, Kexuan Shi, and Shuhang Gu. Uncertainty-guided perturbation for image super-resolution diffusion model. In CVPR, pages 17980–17989, 2025. 6, 2, 3, 5 [53] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586–595, 2018. 6 [54] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3096– 3105, 2019. 1, 2 [55] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu.Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286–301, 2018. 2 AlignVAR: Towards Globally Consistent Visual Autoregression for Image Super-Resolution Supplementary Material This supplementary material provides additional analyses and results that extend the findings in the main paper. To clarify the procedural details of our method, §Sec. 1 presents the complete training and inference pipelines of AlignVAR. To assess computational efficiency, §Sec. 2 pro- vides both theoretical and empirical complexity analyses. To further verify perceptual quality, §Sec. 3 reports the re- sults of a user study comparing AlignVAR with several state-of-the-art baselines. To evaluate real-world robust- ness, §Sec. 4 includes additional experiments on the chal- lenging RealLR200 dataset. To better understand the be- havior of our model, §Sec. 5 visualizes multi-scale attention maps produced by AlignVAR. To examine the contribution of each proposed component, §Sec. 6 conducts an ablation analysis of SCA and HCC using the Edge IoU metric. To investigate how SCA affects the receptive range, §Sec. 7 compares attention maps with and without SCA. To ana- lyze the influence of the hyperparameter λ, §Sec. 8 stud- ies its effect on loss convergence and prediction accuracy across scales. To further showcase visual improvements, §Sec. 9 presents extended qualitative comparisons. Finally, §Sec. 10 discusses the limitations of AlignVAR and outlines possible directions for future work. 1. Training and Inference Procedures This section provides the procedural details of AlignVAR during both training and inference. All notations follow those introduced in the main paper. The overall training pipeline is summarized in Algorithm 1, and the autoregres- sive reconstruction process is presented in Algorithm 2. Training procedure. During training, AlignVAR follows a scale-wise teacher-forcing strategy, as outlined in Algo- rithm 1. Given an HR–LR pair, the frozen VAE encoder [33] extracts the full-resolution latent z ∈R C×H K ×W K , where C denotes the channel dimension and (H K ,W K ) is the resolution of the last scale. For each scale k, the corre- sponding ground-truth latent u gt k ∈ 1,...,|V| H k ×W k is obtained via spatial downsampling and quantization, where (H k ,W k ) denote the height and width of scale k, and |V| denotes the size of the codebook. The residual token is de- fined as r k gt = u k gt − u k−1 gt . The autoregressive predictor p θ receives the reweighted context ̃r 1:k−1 gt , where the mod- ulation mask m k ∈R 1×H k ×W k is produced by the mask generator M φ using the structural guidance extracted from the LR input. The model predicts the categorical distribu- Algorithm 1 AlignVAR Training 1: Inputs: HR–LR image pair (I HR ,I LR ) 2: Hyperparams: steps K, resolutions(H k ,W k ) K k=1 3: z =E (I HR ); c =E con (I LR ); s =|Laplacian(I LR )|; 4: u 0 gt = 0, u 0 pred = 0; 5: for k = 1,...,K do 6: u k gt = Quant(Down(z,H k ,W k )); 7: r k gt = u k gt − u k−1 gt ; 8: end for 9: for k = 1,...,K do 10: ̄s k = norm(Down k (s)); 11: m k = σ(M φ ([r k gt , ̄s k ])); 12: ̃r k gt = (1 + m k )⊙ r k gt ; 13:Predict p θ ( ̃r k | ̃r 1:k−1 gt ,c); 14: u k pred = u k−1 pred + ˆr k pred ; 15: end for 16: ComputeL CE andL HCC ; 17: Update (θ,φ); 18: return trained model parameters; tion over the |V| codebook entries for r k , and the param- eters of p θ and M φ are jointly optimized using the multi- scale cross-entropy loss L CE and the hierarchical consis- tency lossL HCC applied to the cumulative predictions. Inference procedure. During inference, the model per- forms fully autoregressive multi-scale prediction, as de- tailed in Algorithm 2. The LR input is first encoded into the conditional latentc∈R C×H K ×W K , and the structural guid- ance map is computed in the same way as during training. Starting from u 0 pred = 0, the model sequentially predicts the residual tokens ˆr k ∈1,...,|V| H k ×W k , generates the mask m k , computes the reweighted token map ̃r k , and up- dates the cumulative latent state u k pred . After all scales are completed, the final latent is mapped through the shared VQ codebook and decoded by the VAE decoder to produce the super-resolved output. 2. Complexity Analysis Theoretical complexity. We analyze the computational cost of AlignVAR by examining its multi-scale autore- gressive generation process.Let the latent resolution sequence be (h 1 ,w 1 ), (h 2 ,w 2 ),..., (h K ,w K ), where (h k ,w k ) denotes the height and width of the VQ code map Algorithm 2 AlignVAR Inference 1: Inputs: LR image I LR 2: Hyperparams: steps K, resolutions(H k ,W k ) K k=1 3: c =E con (I LR ); s =|Laplacian(I LR )|; 4: u 0 pred = 0; 5: for k = 1,...,K do 6: ̄s k = norm(Down k (s)); 7:ˆr k = arg maxp θ ( ̃r k | ̃r 1:k−1 ,c); 8: m k = σ(M φ ([ ˆr k , ̄s k ])); 9: ̃r k = (1 + m k )⊙ ˆr k ; 10: u k pred = u k−1 pred + ˆr k ; 11: end for 12: f K = lookup(V, u K pred ); 13: I SR = D(f K ); 14: return reconstructed SR image I SR ; at the k-th autoregressive step, and the final resolution sat- isfies h K = h, w K = w. For simplicity, we assume n k = h k = w k for all scales. Following the progressive- resolution design [31], we set n k = a k−1 with a constant a > 1, chosen such that a K−1 = n. At scale k, the model attends to all accumulated token maps (r 1 ,r 2 ,...,r k ), and the total number of tokens is k X i=1 n 2 i = k X i=1 a 2(i−1) = a 2k − 1 a 2 − 1 .(14) Thus, the computational cost of the k-th autoregressive step is a 2k − 1 a 2 − 1 2 .(15) Summing over all steps yields the total generation com- plexity: log a (n)+1 X k=1 a 2k − 1 a 2 − 1 ! 2 = (a 4 − 1) logn + a 8 n 4 − 2a 6 n 2 − 2a 4 (n 2 − 1) + 2a 2 − 1 loga (a 2 − 1) 3 (a 2 + 1) loga ∼O(n 4 ). (16) Since the final resolution satisfies a K−1 = n, the to- tal complexity is dominated by the last autoregressive step, yielding an overall complexity of O(n 4 ). Next, we consider the additional components introduced by AlignVAR, including the structural-guidance mask gen- erator M φ and the residual modulation operation. The mask map at scale k is denoted by m k ∈R 1×n k ×n k . Comput- ing m k requires a single forward pass over the feature map, followed by element-wise modulation, both of which scale linearly with n 2 k . Since this is asymptotically smaller than the autoregressive cost P k i=1 n 2 i 2 , the additional mod- ules do not influence the overall complexity. 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. Align- VAR achieves high performance with substantially lower com- plexity than other models. In summary, the theoretical generation complexity of AlignVAR remains O(n 4 ), showing that the proposed consistency-enhancing modules maintain the same asymp- totic computational cost. Empirical comparison. To further evaluate the practi- cal complexity, we compare AlignVAR with representa- tive diffusion-based SR methods and the VAR-based base- line VARSR. We measure inference time, FLOPs, and per- ceptual quality (CLIP-IQA [34]) under the same hardware and input configuration. As shown in Fig. 11, AlignVAR achieves the best accuracy–efficiency balance: it delivers significantly higher CLIP-IQA than diffusion models while maintaining much lower FLOPs and substantially faster in- ference. Compared with VARSR, AlignVAR attains a large perceptual gain while introducing negligible computational overhead, consistent with our theoretical analysis. These re- sults demonstrate that AlignVAR retains the computational advantages of VAR while notably enhancing global consis- tency and reconstruction quality. 3. User Study To evaluate the perceptual quality of the super-resolution results generated by AlignVAR compared to other meth- ods, we conducted a user study. We randomly selected 50 images from three different datasets: RealSR [36], DRealSR [43], and DIV2K-val [3]. These images were then presented to 20 participants, who were asked to compare the super-resolution results produced by AlignVAR against those from several competing models. The models evalu- ated in the study included VARSR [28], PASD [46], Diff- BIR [23], UPSR [52], StableSR [35], and LDM [29]. Par- ticipants were asked to choose which model’s output they preferred for each image. The results of the user study are shown in Fig. 12, which highlight the superiority of 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. DatasetsMetrics GAN-basedDiffusion-basedVAR-based BSRGANReal-ESRSwinIRLDMStableSRDiffBIRPASDUPSRVARSRAlignVAR RealLR200 NIQE↓4.38174.20484.21574.25334.25164.17154.93304.76064.45794.0617 MANIQA↑0.54620.55820.37410.30490.58410.60660.59020.42060.45360.6237 CLIPIQA↑0.56790.53890.55960.52530.60680.67970.65090.63970.61440.6734 MUSIQ↑64.8762.9463.5555.1963.3068.2062.0666.4662.1269.36 42.87%42.87% OursUPSRPSAD DiffBIR VARSR LDMStableSR Figure 12. User study results comparing the super-resolution out- puts of different models. AlignVAR in producing high-quality reconstructed images as judged by human evaluators, further validating the effec- tiveness of the proposed method. 4. Evaluation on the RealLR200 Dataset To further assess the robustness and real-world applicabil- ity of AlignVAR, we conduct additional experiments on the RealLR200 [44] dataset, a challenging collection of real-world low-resolution images with diverse and complex degradations. We compare AlignVAR against representa- tive GAN-based methods (BSRGAN [51], Real-ESR [39], SwinIR [21]), diffusion-based approaches (LDM [29], Sta- bleSR [35], DiffBIR [23], PASD [46], UPSR [52]), and the VAR-based baseline VARSR [28]. Both quantitative and qualitative comparisons are performed. Table 6 summarizes the no-reference quality evaluation across four widely used perceptual metrics: NIQE [26], MANIQA [45], CLIP-IQA [34], and MUSIQ [14]. Align- VAR consistently achieves the best or second-best perfor- mance among all competing methods. Notably, AlignVAR surpasses the diffusion-based models, which typically excel in perceptual realism, demonstrating a superior balance be- tween fidelity, sharpness, and naturalness. Compared with VARSR [28], AlignVAR obtains substantial improvements across all metrics, highlighting the effectiveness of the pro- posed consistency-enhancing mechanisms. Qualitative results on RealLR200 are presented in Fig. 17. GAN-based models introduce unnatural textures or hallucinated structures, while diffusion-based methods may produce inconsistent details under strong degradations. In contrast, AlignVAR reconstructs coherent edges, stable tex- tures, and visually pleasing details. The improvements are evident in object boundaries and repeated patterns, where spatial and hierarchical consistency play a crucial role. 5. Attention Map Visualization To further demonstrate the effectiveness of our method in expanding the attention range and enhancing the reweight- ing process, we visualize the attention maps at various scales. Specifically, we display the attention maps for all scales except for the first one, as shown in Fig. 13. From the visualizations, it is evident that our approach does not limit the attention to the diagonal elements but instead ex- tends the attention across the entire spatial domain at all scales. This highlights the ability of our model to effectively reweight and capture dependencies over larger regions as the resolution increases. The expansive attention observed at higher scales is a clear indication that the reweighting mechanism in AlignVAR facilitates a more global under- standing of the image structure. 6. Ablation Study on SCA and HCC We perform an ablation study to assess the individual and combined impact of SCA and HCC on the performance of AlignVAR. Specifically, we evaluate the Edge Intersection over Union (Edge IoU) [47] across different scales, com- puted between the predicted edge map and the ground-truth edge map. The edge maps are extracted using the Canny edge detector, and the IoU is then calculated on the result- ing binary edge masks to quantify the accuracy of struc- tural recovery. Fig. 14 presents the results of the ablation study. The full AlignVAR model, which incorporates both SCA and HCC, consistently achieves the highest Edge IoU. As the scale increases, the benefit becomes more apparent, demonstrating the importance of enforcing multi-scale con- Scale 2×2 Scale 3×3 Scale 4×4 Scale 6×6Scale 9×9Scale 13×13 Scale 32×32Scale 24×24Scale 18×18 Scale 32×32Scale 24×24Scale 18×18 Scale 32×32Scale 24×24Scale 18×18 Figure 13. Attention maps at different scales. AlignVAR expands the attention range, capturing dependencies over large areas. sistency in high-resolution reconstruction. When compar- ing the variants, removing SCA leads to the most signifi- cant performance drop. This indicates that SCA plays a pri- mary role in expanding the spatial attention range and main- taining spatial consistency, which is essential for recover- ing fine-grained and structurally aligned edges. Removing HCC also results in a decline in Edge IoU, though the degra- dation is less severe. This suggests that HCC mainly con- tributes to stabilizing cross-scale dependencies, but its in- fluence on edge localization is comparatively weaker than that of SCA. 7. Impact of SCA on Attention Range To assess the influence of the SCA on the attention range, we conduct an ablation study by comparing the attention maps with and without the inclusion of SCA. The results are presented in Fig. 15. Without SCA, the attention maps at different scales demonstrate a more localized focus, with attention primarily concentrated along the diagonal. This suggests that, in the absence of SCA, the model tends to focus on nearby regions, limiting its ability to capture long- range dependencies across the spatial domain. In contrast, when SCA is incorporated, the attention maps exhibit a no- ticeable expansion. The model begins to attend to a much broader area beyond the diagonal, with significant atten- tion given to regions that are further apart. This compari- son demonstrates that SCA significantly increases the atten- tion range, allowing the model to better capture both local and global dependencies, which contributes to the improved performance of AlignVAR. 5678910 Scales 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Edge IOU VARSR AlignVAR w/o HCC AlignVAR w/o SCA AlignVAR 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 hierar- chical consistency, but its influence is comparatively weaker. Scale 2×2 Scale 3×3 Scale 4×4 Scale 6×6Scale 9×9Scale 13×13 Scale 32×32Scale 24×24Scale 18×18 Scale 32×32Scale 24×24Scale 18×18 Scale 32×32Scale 24×24Scale 18×18 w/o SCA with SCA Figure 15. Impact of Spatial Consistency Autoregression (SCA) on attention maps. The attention range is significantly broader when SCA is applied. 8. Effect of Hyperparameter We investigate the effect of the hyperparameter λ on the model’s performance. Specifically, we analyze the impact of four different values of λ: 0.5, 1.0, 1.5, and 2.0, on the loss and accuracy across different scales. The results are shown in Fig. 16. As observed, when λ is small (e.g., 0.5), the model converges faster, but the token prediction accuracy is lower, particularly for the middle scales such as 6× 6 and 9× 9. In contrast, when λ is set to 1.0, the model achieves the highest accuracy, with a more fast convergence. Asλ increases further (1.5 and 2.0), the loss converges more slowly, and the accuracy starts to decrease, suggesting that higher values of λ may hinder the model’s ability to reach optimal performance for these scales. In summary, the value of λ plays a crucial role in balancing convergence speed and accuracy. A λ value of 1.0 provides the best trade-off be- tween fast convergence and high accuracy. Figure 16. Effect of the hyperparameter λ on training dynamics. We compare four values of λ (0.5, 1.0, 1.5, 2.0) in terms of loss and token prediction accuracy across different scales. A smaller λ leads to faster convergence but lower accuracy, particularly at the middle scales (e.g., 6× 6 and 9× 9), while λ = 1.0 achieves the best overall accuracy with stable convergence. Larger λ values slow down loss convergence and degrade accuracy, indicating that excessively large consistency constraints hinder optimal learning. 9. Additional Visual Results In this section, we present additional visual results to demonstrate the effectiveness of AlignVAR across various images. The results in Fig. 18 and Fig. 19 highlight the recovery of fine details in different scenes, with compar- isons between AlignVAR and other state-of-the-art mod- els, including GAN-based models (Real-ESRGAN [39], BSRGAN [51]) and diffusion-based models (PASD [46], UPSR [52]). In Fig. 18, AlignVAR performs notably better in recover- ing intricate details such as the beak of the penguin, archi- tectural structures. GAN-based models like BSRGAN, in contrast, tend to generate overly smooth results, losing fine LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0804DIV2K-0804 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0812DIV2K-0812 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0801DIV2K-0801 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0836DIV2K-0836 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0846DIV2K-0846 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0830DIV2K-0830 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0862DIV2K-0862 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0847DIV2K-0847 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVAR103103 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVAR042042 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVAR185185 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVAR135135 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. LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0804DIV2K-0804 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0812DIV2K-0812 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0801DIV2K-0801 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0836DIV2K-0836 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0846DIV2K-0846 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0830DIV2K-0830 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0862DIV2K-0862 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0847DIV2K-0847 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. LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0804DIV2K-0804 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0812DIV2K-0812 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0801DIV2K-0801 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0836DIV2K-0836 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0846DIV2K-0846 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARNikon-046Nikon-046 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0862DIV2K-0862 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVARDIV2K-0844DIV2K-0844 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVAR103103 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVAR042042 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVAR185185 LRLR HRHR Real-ESRGANReal-ESRGANBSRGANBSRGANPASDPASD UPSRUPSRVARSRVARSRAlignVARAlignVAR135135 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. details, especially in intricate structures like the penguin’s beak or the architectural details. Diffusion models such as PASD also struggle with fine texture recovery, often result- ing in overly blurred outputs, especially in high-frequency areas. However, compared to the HR image, some gaps remain, especially in areas with complex texture. This is likely due to the difficulty in recovering fine details, such as small textures in challenging low-resolution inputs. In Fig. 19, AlignVAR excels in recovering sketch-like images and animal details, such as the eye and fur texture of the animal. The eyes are restored with sharpness, natural shine, and detailed textures that are closer to the ground truth com- pared to other methods. GAN models like BSRGAN tend to oversmooth such textures, leading to unnatural results, while diffusion models again introduce unnecessary blur- ring, especially in small, detailed areas like the eyes or fur. 10. Limitation While AlignVAR achieves strong performance across a wide range of benchmarks, several limitations remain. First, the recovery of middle scales proves to be particularly chal- lenging, as shown in our hyperparameter analysis 8. These intermediate resolutions more sensitive to hyperparameter choices such as λ. In future work, we plan to explore dy- namically adjustable consistency constraints that adapt to the difficulty of each scale. Second, although AlignVAR improves the spatial and hierarchical consistency of autore- gressive prediction, its computation is still dominated by the final high-resolution scale, inheriting the O(n 4 ) com- plexity of VAR-based models. This may limit scalability to extremely high-resolution image generation. Finally, our method relies on the quality of VQ tokenization. When the codebook fails to accurately represent rare textures or sub- tle structures, even consistent autoregression cannot fully recover the missing details. Incorporating stronger gener- ative priors or adaptive codebook refinement may further enhance robustness to challenging degradations.