Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Detail Continuation over a Trustworthy Coarse Scale for Autoregressive Super-Resolution

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that shortening the autoregressive path of visual autoregressive super-resolution by generating the coarse scales directly from the low-resolution input reduces hallucination without sacrificing fidelity.

desk verdict K2N is a plausible, incremental improvement over VARSR with a real evaluation gap: the 'trustworthy coarse scale' premise is never directly measured. read the letter →

arxiv 2608.01823 v2 pith:ICMTEHWA submitted 2026-08-03 cs.CV

Hongyi Fang , Jiahui Wu , Yichen Yue , Benjia Zhou , Dan Zeng This is my paper
classification cs.CV
keywords imagesuper-resolutionvisualautoregressivemodelinghallucinationmitigationcoarse-to-finegenerationnext-scalepredictiondetailcontinuationblind
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that generative super-resolution should not regenerate the coarse scales that the low-resolution input already constrains. It proposes K2N, a modification of visual autoregressive (VAR) super-resolution that replaces the first M=3 autoregressive scale predictions with one parallel feed-forward estimate of coarse-scale residual features taken directly from the LR image, then continues the remaining fine scales autoregressively with a frozen pretrained transformer. The central claim is that this k-to-N detail continuation reduces accumulated autoregressive error and hallucinated details while keeping standard fidelity metrics broadly competitive. If true, it matters because it gives a concrete, scale-wise way to separate trustworthy structure from uncertain detail in generative restoration, and it does so at lower inference cost.

What carries the argument

The machinery is the coarse-scale prefix reconstructor: a module that maps LR to a stack of continuous residual features $z_k = f - \sum_{m=1}^{k-1} \mathrm{upsample}(\mathrm{lookup}(V, r_m))$, i.e., the residual information left after earlier scales. By predicting these residuals in parallel instead of token-by-token, then applying the same scale-wise accumulation operator $A(\cdot)$ used by VAR to build transformer prefixes, K2N injects LR evidence directly into the generation context. The pretrained VAR transformer is untouched, so the only learned component is this prefix predictor and its three-loss supervision.

What would settle it

A direct test is to run K2N with M ranging from 1 to K: if hallucination scores keep improving as more scales are prefilled, or if the predicted coarse prefixes are no closer to the HR residuals than the original autoregressive states are, then the load-bearing premise fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that the full 1-to-N next-scale autoregressive path used by VAR-based super-resolution is unnecessary for early scales. K2N predicts the first M continuous residual features $\hat{z}_1, \dots, \hat{z}_M$ from the LR input in parallel, accumulates them into prefix states $\hat{s}_k = A(\hat{z}_{1:k})$, and prefills these into the frozen VAR transformer so that only scales $M+1$ through $K$ are generated autoregressively. Training supervises the predictor at three levels: L1 on residuals, codebook-aligned cross-entropy to stay compatible with the pretrained codebook, and L1 on accumulated prefix states. With $M=3$, K2N matches VARSR closely on PSNR, SSIM, LPIPS, and DISTS while improving no-reference perceptual metrics (CLIPIQA, MUSIQ, MANIQA) on all three benchmarks and scoring highest on the paper's hallucination-focused MLLM evaluation.

Load-bearing premise

The load-bearing premise is that the coarse structural information of the high-resolution image is still present and recoverable from the low-resolution input, so it can be filled in directly instead of guessed step by step, and that errors from this direct fill-in do less harm than errors accumulated through the original autoregressive chain.

Editorial extensions

If this is right

  • If K2N is right, autoregressive SR no longer needs to generate every scale from scratch: coarse scales can be prefilled from LR, shortening the effective chain and reducing error accumulation.
  • K2N keeps the pretrained VAR transformer frozen, so the gains come from a small added prefix predictor plus fewer generation steps (7 vs 10 for VARSR), not from retraining the backbone.
  • The three-level supervision—L1 on residuals, codebook-aligned cross-entropy, and L1 on accumulated prefix states—shows how to inject LR evidence into a pretrained next-scale model.
  • On the paper's 100-image hallucination-focused set, K2N scores above VARSR, SeeSR, and FaithDiff, and users prefer it in pairwise comparisons.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural but untested extension is a scale-adaptive M: instead of fixing M=3, stop prefilling at the scale where LR evidence stops being reliable, which the paper's own failure cases suggest may lie beyond the earliest scales.
  • The 100-image diagnostic set uses a multimodal LLM as judge; a larger benchmark of LR-HR pairs with identifiable structures would test whether the hallucination advantage holds beyond this set.
  • The same k-to-N continuation idea should transfer to other next-scale restoration tasks, such as deblurring or deraining, wherever the input preserves low-frequency structure.
  • Because K2N only adds a small parallel module to a frozen transformer, the recipe could be applied to future, larger VAR backbones without retraining them.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes K2N, a modification of visual autoregressive (VAR) super-resolution (SR) that replaces the autoregressive generation of the first M=3 coarse scales with a parallel predictor conditioned on the low-resolution (LR) input, before continuing with the pretrained VAR transformer for the remaining finer scales. The authors argue that coarse-scale information in LR is relatively reliable and that a deterministic parallel prefix avoids error accumulation in a long autoregressive chain. Experiments on DIV2K-Val, RealSR, and DRealSR compare K2N against nine baselines, and a newly constructed 100-image hallucination-focused set is used with MLLM-based scoring and a user study to claim improved hallucination mitigation.

Significance. If the central premise holds, K2N is a simple, efficient, and practical reformulation of VAR-based SR: it reduces the autoregressive path, maintains competitive standard metrics, and improves faithfulness to the LR input. The paper ships a broad benchmark comparison (9 baselines, 3 datasets), reports inference efficiency gains, and commits to code release. However, the paper's titular 'trustworthy coarse scale' is not directly measured, and the hallucination evaluation is deferred to a supplementary document, so the significance of the claimed contribution cannot yet be fully assessed.

major comments (4)
  1. [§3.3 and Figure 2] The central premise that coarse-scale residual features can be reliably predicted from LR is not validated. Figure 2 reports L1 distances between degraded and HR residual features, which characterizes the degradation, not the fidelity of the learned predictor. The paper never reports the inference-time error of the predicted prefix states z_hat_k or s_hat_k against HR-derived ground truth, nor does it compare this error with the autoregressive prefix error of VARSR. Without such a measurement, the hallucination reduction in Table 2 could be attributed to the shortened autoregressive path or the specific choice of M=3 rather than to the trustworthiness of the predicted coarse scale. Please add a quantitative comparison of prefix-state fidelity between K2N and VARSR.
  2. [§4.1, Table 4] The choice of M=3 is fixed and never ablated. Since the method's motivation is scale-dependent trustworthiness, an ablation over M (e.g., M=1,2,3,4,5) reporting both standard metrics and hallucination scores is necessary to support the claim that the benefit comes from the parallel coarse prefix. The absence of this ablation also prevents understanding the trade-off between the amount of deterministic prefix and the generative flexibility of the autoregressive continuation.
  3. [§4.3 and Table 2] The hallucination-focused evaluation is the primary evidence for the paper's main contribution, but it is not self-contained in the reviewed version. The construction of the 100-image set, the screening thresholds, the exact MLLM prompt, and the scoring instructions are only referenced to a supplementary document. For a reviewable and reproducible claim, the full protocol (or at least a precise description of the set and prompt) must be included in the main text or an available appendix.
  4. [§4.2 and Table 1] The statement that K2N 'remains competitive' with VARSR on standard SR metrics is partly based on PSNR drops that are not negligible: 23.99 to 23.81 on DIV2K-Val, 25.57 to 24.98 on RealSR, and 28.15 to 27.18 on DRealSR. Since the paper claims hallucination mitigation while preserving fidelity, the authors should explicitly discuss these drops and show that they are within an acceptable fidelity budget, or temper the claim accordingly.
minor comments (5)
  1. [§3.3] There is a duplicated sentence: 'Motivated by this observation, we no longer generate the first M coarse scales autoregressively.' appears twice in the same paragraph. Please remove the duplicate.
  2. [§4.1] The phrase 'Unless otherwise specified.' ends a sentence but is followed by a period and then a new sentence starting with 'For evaluation' without a space; this appears to be a punctuation/formatting error. Please fix the spacing and punctuation.
  3. [Figure 2] The caption does not state which dataset or degradation types are used for the L1-distance plot, nor how the distances are aggregated (e.g., mean over images, standard deviation). Please specify the experimental conditions and show error bars or variance information.
  4. [Table 3] The ablation of prefix prediction forms reports only DIV2K-Val standard metrics and perceptual metrics; it does not report hallucination scores. Since the method's contribution concerns hallucination mitigation, the ablation should include the hallucination-focused evaluation.
  5. [§5] The failure-case discussion mentions 'reliable LR evidence may also persist at intermediate scales currently modeled autoregressively.' This is an interesting observation that could also be used to motivate an adaptive M; please consider connecting it to the missing M-ablation when revising.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: K2N's derivation is self-contained and its central claim is not equivalent to its inputs by construction.

full rationale

The paper's derivation chain was walked from Eq. 1 (the standard VAR full-path factorization) through Eq. 3 (replacing the first M coarse-scale terms with a learned parallel predictor), Eq. 5 (continuing the remaining fine scales with the frozen VAR transformer), and Algorithm 1, with losses L_res, L_cb, and L_acc (Eqs. 6-9) supervising the predictor against HR-derived ground-truth residual, codebook, and accumulated-prefix states. At inference, the predicted prefix is prefilled and the remaining scales are generated autoregressively, so the output is a genuine composition of a learned prefix estimator and the external VARSR backbone. The evaluation compares against external baselines on standard benchmarks (DIV2K-Val, RealSR, DRealSR) and uses the external Hallucination Score protocol of Ren et al. There is no equation that reduces the reported hallucination improvement to the training objective or to a fitted parameter by construction; the authors' own failure cases (Figure 6) show that fine-detail hallucinations persist, which further indicates the evaluation is not a tautology. Reference [42], cited for error accumulation in autoregressive generation, is not authored by the present authors, and no load-bearing argument rests on a self-citation chain. The main weaknesses are empirical rather than circular: the 'trustworthy coarse scale' premise is motivated by Figure 2, which reports L1 distances between degraded and HR residual features rather than the learned predictor's inference-time prefix error, and the choice M=3 is not ablated. These are validation gaps, not circular reductions, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The ledger captures the main hand-chosen design parameters (M, lambda, beta) and the key domain assumptions about LR reliability and the MLLM judge. No new physical or ontological entities are introduced.

free parameters (3)
  • M (coarse prefix length) = 3
    Number of earliest scale states predicted in parallel instead of autoregressively; set in Section 4.1 without ablation across M.
  • lambda (codebook-aligned loss weight) = 0.001
    Chosen by hand in Section 4.1; no sensitivity analysis is reported.
  • beta (accumulated-state loss weight) = 0.2
    Chosen by hand in Section 4.1; no sensitivity analysis is reported.
assumptions (3)
  • domain assumption The pretrained VARSR model and its scale-wise residual tokenization (Eq. 2) are taken as a fixed, reliable base for fine-scale generation.
    K2N freezes the VAR transformer and only adds a prefix predictor; if the base model's fine-scale behavior were flawed, K2N would inherit it. See Section 3.2 and Algorithm 1.
  • domain assumption LR preserves reliable coarse-scale information under the degradations used, so a deterministic predictor can substitute for autoregressive sampling at early scales.
    Supported only by Figure 2, which shows residual distances without dataset or error statistics; used to justify Eq. 3 in Section 3.3.
  • domain assumption MLLM-based hallucination scores and the self-constructed benchmark measure the intended hallucination concept.
    The protocol is deferred to supplementary; no validation of the judge against human labels is shown in the main text. See Section 4.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Detail Continuation over a Trustworthy Coarse Scale for Autoregressive Super-Resolution." pith.science (2026). https://pith.science/paper/ICMTEHWA

@misc{pith2026260801823,
  author       = {Pith},
  title        = {Pith review of: Detail Continuation over a Trustworthy Coarse Scale for Autoregressive Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ICMTEHWA}},
  note         = {Machine review of arXiv:2608.01823}
}
abstract

Hallucination remains a persistent challenge in generative super-resolution (GSR), where reconstructed results may contain visually plausible yet weakly supported content, structural deviations, or unnatural textures with respect to the low-resolution (LR) input. Existing GSR methods have extensively explored the trade-off between perceptual realism and reconstruction fidelity, but the division between preserving reliable coarse-scale information and restoring more uncertain fine details is often handled implicitly within the overall restoration process. Visual autoregressive (VAR) modeling provides a natural opportunity to revisit this issue, as its coarse-to-fine next-scale prediction offers an explicit scale-wise generation interface. However, existing VAR-based SR methods still inherit the original full 1-to-$N$ autoregressive generation path, even though, for super-resolution, coarse-scale information in LR is often relatively more reliable, while long autoregressive chains may accumulate prediction errors. Motivated by these observations, we propose \textbf{K2N}, which reformulates VAR-based SR from full-path generation into a $k$-to-$N$ detail continuation process. Specifically, early coarse-scale states are established directly from LR, while only the remaining finer scales are restored autoregressively. Experimental results show that K2N remains competitive with the VARSR baseline on standard SR metrics, while exhibiting clearer advantages on hallucination-focused evaluation. These findings suggest that explicitly rethinking the generation path in a scale-wise manner can be a promising direction for improving the reliability of generative super-resolution. Our code will be released soon at https://github.com/BRL-SYSU/K2NSR.

Figures

Figures reproduced from arXiv: 2608.01823 by the authors.

Figure 1
Figure 1. Visual example of hallucination mitigation. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. L1 distance between degraded and HR residual fea [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of K2N during training and inference. K2N reformulates VAR-based super-resolution from full-path [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of different super-resolution methods. K2N restores sharper and more natural details while [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: User study results and blind pairwise comparison examples on the hallucination-focused benchmark. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 28 canonical work pages

  1. [1]

    Eirikur Agustsson and Radu Timofte. 2017. NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

    work page 2017

  2. [2]

    Yochai Blau and Tomer Michaeli. 2018. The Perception-Distortion Tradeoff. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2018

  3. [3]

    Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. 2019. To- ward Real-World Single Image Super-Resolution: A New Benchmark and a New Model. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    work page 2019

  4. [4]

    Junyang Chen, Jinshan Pan, and Jiangxin Dong. 2025. FaithDiff: Unleashing Dif- fusion Priors for Faithful Image Super-resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 28188–28197

    work page 2025

  5. [5]

    Simoncelli

    Keyan Ding, Kede Ma, Shiqi Wang, and Eero P. Simoncelli. 2022. Image Quality Assessment: Unifying Structure and Texture Similarity.IEEE Transactions on 8 Detail Continuation over a Trustworthy Coarse Scale for Autoregressive Super-Resolution Pattern Analysis and Machine Intelligence44, 5 (2022), 2567–2581. doi:10.1109/ TPAMI.2020.3045810

    arXiv 2022

  6. [6]

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. 2016. Image Super-Resolution Using Deep Convolutional Networks.IEEE Transactions on Pattern Analysis and Machine Intelligence38, 2 (2016), 295–307. doi:10.1109/ TPAMI.2015.2439281

    arXiv 2016

  7. [7]

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. 2025. Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 15733–15744

    work page 2025

  8. [8]

    Jonathan Ho and Tim Salimans. 2022. Classifier-Free Diffusion Guidance. arXiv:2207.12598 [cs.LG] https://arxiv.org/abs/2207.12598

    arXiv 2022

Show all 46 references
  1. [9]

    Chenggong Hu, Shaoyin Ma, Yi Wang, Li Sun, Mingli Song, and Jie Song. 2026. Geometry-guided Emotion Modulation for Controllable and Photorealistic Emo- tional Talking Face Generation. arXiv:2608.00663 [cs.CV] doi:10.48550/arXiv. 2608.00663

    2026 doi arXiv

  2. [10]

    Chenggong Hu, Yi Wang, Mengqi Xue, Haofei Zhang, Jie Song, and Li Sun

  3. [11]

    Siyu Jiao, Gengwei Zhang, Yinlong Qian, Jiancheng Huang, Yao Zhao, Humphrey Shi, Lin Ma, Yunchao Wei, and Zequn Jie. 2025. Flexvar: Flexible visual autore- gressive modeling without residual prediction.arXiv preprint arXiv:2502.20313 (2025)

    2025

  4. [12]

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. 2021. MUSIQ: Multi-Scale Image Quality Transformer. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV). 5148–5157

    2021

  5. [13]

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. 2021. SwinIR: Image Restoration Using Swin Transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. 1833–1844

    2021

  6. [14]

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee

  7. [15]

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. 2025. DiffBIR: Toward Blind Image Restoration with Generative Diffusion Prior. InComputer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, ...

    2025

  8. [16]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101(2017)

    2017 arXiv

  9. [17]

    Com- pletely Blind

    Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. 2013. Making a “Com- pletely Blind” Image Quality Analyzer.IEEE Signal Processing Letters20, 3 (2013), 209–212. doi:10.1109/LSP.2012.2227726

    2013

  10. [18]

    Yunpeng Qu, Kun Yuan, Jinhua Hao, Kai Zhao, Qizhi Xie, Ming Sun, and Chao Zhou. 2025. Visual autoregressive modeling for image super-resolution.arXiv preprint arXiv:2501.18993(2025)

    2025 arXiv

  11. [19]

    Yunpeng Qu, Kun Yuan, Kai Zhao, Qizhi Xie, Jinhua Hao, Ming Sun, and Chao Zhou. 2025. XPSR: Cross-Modal Priors for Diffusion-Based Image Super- Resolution. InComputer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol...

    2025

  12. [20]

    Weiming Ren, Raghav Goyal, Zhiming Hu, Tristan Ty Aumentado-Armstrong, Iqbal Mohomed, and Alex Levinshtein. 2025. Hallucination Score: Towards Mitigating Hallucinations in Generative Image Super-Resolution.arXiv preprint arXiv:2507.14367(2025)

    2025

  13. [21]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis With Latent Diffusion Mod- els. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10684–10695

    2022

  14. [22]

    Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. 2024. Hart: Efficient visual gen- eration with hybrid autoregressive transformer.arXiv preprint arXiv:2410.10812 (2024)

    2024 arXiv

  15. [23]

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. 2024. Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction. InAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, ...

    2024 doi

  16. [24]

    Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, koray kavukcuoglu, Oriol Vinyals, and Alex Graves. 2016. Conditional Image Generation with PixelCNN Decoders. InAdvances in Neural Information Processing Systems, D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett (E...

    2016

  17. [25]

    Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. 2016. Pixel Recurrent Neural Networks. InProceedings of The 33rd International Confer- ence on Machine Learning (Proceedings of Machine Learning Research, Vol. 48), Maria Florina Balcan and Kilian Q. Weinberger (Eds....

    2016

  18. [26]

    Chan, and Chen Change Loy

    Jianyi Wang, Kelvin C.K. Chan, and Chen Change Loy. 2023. Exploring CLIP for Assessing the Look and Feel of Images.Proceedings of the AAAI Conference on Artificial Intelligence37, 2 (Jun. 2023), 2555–2563. doi:10.1609/aaai.v37i2.25353

    2023 doi

  19. [27]

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C. K. Chan, and Chen Change Loy. 2024. Exploiting Diffusion Prior for Real-World Image Super- Resolution. arXiv:2305.07015 [cs.CV] https://arxiv.org/abs/2305.07015

    2024 arXiv

  20. [28]

    Siyang Wang, Naishan Zheng, Jie Huang, and Feng Zhao. 2025. Navigating Image Restoration with VAR’s Distribution Alignment Prior. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 7559–7569

    2025

  21. [29]

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. 2021. Real-ESRGAN: Training Real-World Blind Super-Resolution With Pure Synthetic Data. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. 1905–1914

    2021

  22. [30]

    Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. 2018. Recovering Realistic Texture in Image Super-Resolution by Deep Spatial Feature Transform. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2018

  23. [31]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing13, 4 (2004), 600–612. doi:10.1109/TIP.2003.819861

    2004

  24. [32]

    Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. 2020. Component Divide-and-Conquer for Real-World Image Super-Resolution. InComputer Vision – ECCV 2020, Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm (Eds.). Springer...

    2020

  25. [33]

    Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. 2024. SeeSR: Towards Semantics-Aware Real-World Image Super- Resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 25456–25467

    2024

  26. [34]

    Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. 2022. MANIQA: Multi-Dimension Attention Net- work for No-Reference Image Quality Assessment. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

    2022

  27. [35]

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. 2023. IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models. arXiv:2308.06721 [cs.CV] https://arxiv.org/abs/2308.06721

    2023 arXiv

  28. [36]

    Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. 2022. Scaling autoregressive models for content-rich text-to-image generation.arXiv preprint arXiv:2206.107892, 3 (2022), 5

    2022 arXiv

  29. [37]

    Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. 2021. Designing a Practical Degradation Model for Deep Blind Image Super-Resolution. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 4791–4800

    2021

  30. [38]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding Conditional Control to Text-to-Image Diffusion Models. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV). 3836–3847

    2023

  31. [39]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang

  32. [40]

    Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. 2019. RankSRGAN: Gen- erative Adversarial Networks With Ranker for Image Super-Resolution. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    2019

  33. [41]

    Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. 2018. Image Super-Resolution Using Very Deep Residual Channel Attention Networks. InProceedings of the European Conference on Computer Vision (ECCV)

    2018

  34. [42]

    Gengze Zhou, Chongjian Ge, Hao Tan, Feng Liu, and Yicong Hong. 2025. Re- thinking Training Dynamics in Scale-wise Autoregressive Generation.arXiv preprint arXiv:2512.06421(2025)

    2025

  35. [43]

    Qiwen Zhu, Yanjie Wang, Shilv Cai, Liqun Chen, Jiahuan Zhou, Luxin Yan, Sheng Zhong, and Xu Zou. 2024. Perceptual-Distortion Balanced Image Super- Resolution is a Multi-Objective Optimization Problem. InProceedings of the 32nd ACM International Conference on Multimedia(Melbour...

    2024

  36. [2017]

    In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

    Enhanced Deep Residual Networks for Single Image Super-Resolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

  37. [2018]

    InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

  38. [2026]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    Semi-supervised latent disentangled diffusion model for textile pattern generation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 4798–4806

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.