Pith. sign in

REVIEW 4 major objections 5 minor 44 references

Multi-scale Image Super Resolution with a Single Auto-Regressive Model

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 a single auto-regressive transformer, equipped with a hierarchical tokenizer and a preference-based loss, super-resolves images at multiple scales in one forward pass and matches or beats much larger models.

desk verdict The hierarchical multi-scale tokenizer is a real building block; the DPO loss needs specification and the SOTA claim needs toning down, but the paper deserves a serious review. read the letter →

arxiv 2506.04990 v1 pith:XKJDAI2C submitted 2025-06-05 cs.CV

Enrique Sanchez , Isma Hadji , Adrian Bulat , Christos Tzelepis , Brais Martinez , Georgios Tzimiropoulos This is my paper
classification cs.CV
keywords imagesuper-resolutionvisualautoregressivemodelsnext-scalepredictionhierarchicaltokenizationresidualquantizationdirectpreferenceoptimizationmulti-scalegenerationRQ-VAE
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

Image super-resolution models are usually built for one fixed upscale factor and often rely on huge models or private data. This paper claims that visual auto-regressive (VAR) models, which generate images by predicting residual tokens across increasing scales, are a natural fit for super-resolution provided the tokenizer is rebuilt so that partial token sequences decode into valid intermediate-scale images. The authors introduce a hierarchical RQ-VAE that reuses tokens across scales, and a Direct Preference Optimization (DPO) loss that nudges the transformer to prefer high-resolution token sequences over low-resolution ones. With a 310M-parameter transformer trained only on standard ISR benchmarks, they report state-of-the-art results at 2x and 4x upscaling in a single forward pass, without external training data.

What carries the argument

The load-bearing mechanism is the hierarchical tokenizer (Algorithm 1), which partitions the L=10 residual-quantization steps into three scale groups (s=0.25, 0.5, 1). At each scale, the model tokenizes the downsampled feature map, reusing prior tokens as residuals, and the vocabulary and decoder of a pre-trained RQ-VAE are finetuned so these prefixes reconstruct valid images at 128, 256, and 512. The second component is the DPO-style loss, $-\log\sigma(\beta \log p(z_{HR})/p(z_{LR}))$, where $z_{LR}$ is the tokenization of the bilinearly upsampled low-resolution image; this loss steers the autoregressive transformer away from simply copying the low-resolution structure and toward high-resolution detail, without needing negative samples.

What would settle it

On the DIV2K validation set, compute PSNR and LPIPS between images decoded from the first three residual steps (scale 128) and the corresponding ground-truth images downsampled to 128; if these intermediate reconstructions are not close to the downsampled originals, the hierarchical tokenizer's multi-scale claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the residual quantizer, not the autoregressive head, is what prevents VAR models from producing reliable multi-scale outputs. The proposed Hierarchical Image Tokenization applies residual quantization to progressively larger downsampled versions of the input, forcing higher scales to reuse the tokens from lower scales, so that the first L' residuals decode to a valid image at the corresponding intermediate resolution. After finetuning only the vocabulary and decoder of a pre-trained RQ-VAE, the tokenizer can map any prefix of the token sequence to 128, 256, or 512 resolutions. On top of this, the paper adds a DPO-based regularization term, $-\log\sigma(\beta \log p(z_{HR})/p(z_{LR}))$, which penalizes the model when the predicted tokens are closer to a bilinear upsampling of the low-resolution input. Together these components let a single 310M transformer denoise and super-resolve at 1x, 2x, and 4x in one forward pass, and the paper reports state-of-the-art results on DIV2K-val, RealSR, and DRealSR while using only standard training datasets.

Load-bearing premise

The load-bearing premise is that finetuning only the vocabulary and decoder of a pre-trained RQ-VAE, while keeping the encoder fixed, is enough to make partial token sequences decode into semantically consistent images at 128, 256, and 512.

Editorial extensions

If this is right

  • The same model can output 1x (denoised), 2x, and 4x super-resolutions from a single forward pass, so a deployed system can serve multiple upscale factors without separate models.
  • Because DPO regularization needs only the LR and HR tokenizations, it removes the requirement of collecting negative samples for classifier-free guidance, simplifying the training pipeline.
  • A 310M-parameter transformer trained on standard ISR datasets (DIV2K, DIV8K, Flickr2k, OST, FFHQ subset) is sufficient to be competitive with a 1B model trained on a large private dataset.
  • The hierarchical tokenizer makes partial VAR token sequences decodable, which is also a prerequisite for progressive generation or editing at intermediate resolutions.
  • Intermediate-scale outputs are semantically consistent with the final image, meaning the model's 2x result is a genuine intermediate step rather than an arbitrary partial reconstruction.

Reading between the lines

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

  • A direct extension is to re-partition the residual steps into more than three scales, potentially letting a single model serve arbitrary upscale factors (e.g., 1.5x, 3x) without retraining.
  • The DPO-style loss could transfer to other image-to-image autoregressive tasks where paired degraded and clean inputs exist, such as inpainting, deblurring, or denoising.
  • If the hierarchical tokenizer is a plug-in replacement for standard RQ-VAEs, it might also improve multi-resolution generation in class-to-image or text-to-image VAR models beyond super-resolution.
  • A testable hypothesis is that the intermediate 256 output improves downstream tasks (e.g., object detection) relative to a bilinear-upsampled input; if true, the model would be useful for cost-sensitive pipelines.
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 a multi-scale image super-resolution (ISR) method built on a visual autoregressive (VAR) model. It introduces two components: a hierarchical RQ-VAE tokenizer that aligns token residuals across scales so that partial token sequences can be decoded into valid intermediate-resolution images (128, 256, and 512), and a DPO-style regularization term that encourages the autoregressive transformer to prefer HR token sequences over LR token sequences. The transformer has 310M parameters, is initialized from a VAR d-16 checkpoint, and is trained on standard SR datasets. The method is evaluated on DIV2K-Val, RealSR, and DRealSR at three resolutions, with comparisons to GAN-, diffusion-, and AR-based baselines. The paper claims state-of-the-art ISR performance without external training data and the first use of preference-based optimization for VAR training.

Significance. The hierarchical tokenization idea is genuinely interesting: Table 1a and Figure 2 provide evidence that the finetuned tokenizer can decode intermediate scales with reasonable fidelity, which would enable a single AR model to serve multiple upsampling factors. The DPO ablation in Table 1b shows consistent gains on RealSR and DRealSR across all three resolutions, suggesting that the proposed regularization is practically useful. The use of a 310M model rather than a 1B model is also a practical advantage. However, the paper currently overstates its results: the state-of-the-art claim in the abstract is not supported by Table 3 on DIV2K-Val and RealSR, the DPO objective is under-specified and not reproducible from the text, and the ablation design conflates the two proposed components. These issues are fixable, and the core idea appears defensible, so the paper merits a major revision rather than rejection.

major comments (4)
  1. [Section 3, Eq. (4), and Section 4.1] The DPO regularizer is under-specified. Eq. (4) defines L_DPO = -log sigma(beta log p(zHR)/p(zLR)), and the text defines p(zLR) as the product of softmax probabilities of the LR tokens using the VAR output logits, but it never states which token sequence is fed to the transformer to obtain those logits. If the logits come from the standard teacher-forced forward pass on zHR, then p(zLR) is a per-position probability of LR tokens conditioned on HR history, which is not a valid sequence-level likelihood and makes Eq. (4) a different objective from the stated preference ratio. If a separate forward pass over zLR is required, that pass is not described, and the equal-weight training setup in Section 4.1 does not mention the additional computation. Because Table 1b reports large DPO gains (e.g., RealSR 128 PSNR 20.56 to 22.09), the reported improvements cannot be reproduced without this detail. Please specify the exact conditioning prefix used for scoring zLR and whether a second forward pass is performed.
  2. [Abstract and Table 3] The claim of state-of-the-art results on ISR is not supported by the paper's own benchmark table. On DIV2K-Val, the proposed method has PSNR 24.17, below BSRGAN (24.42) and Real-ESRGAN (24.30), and its LPIPS 0.358 is worse than StableSR (0.3228), SeeSR (0.3283), and VARSR (0.3260). On RealSR, the proposed method has PSNR 25.91, below BSRGAN (26.38), ResShift (26.31), and SwinIR (25.88), and its SSIM 0.756 is below SwinIR (0.7671) and BSRGAN (0.7651). Only on DRealSR does the method achieve the best scores across all metrics. The abstract and Section 4.3 should be revised to qualify the claim, for example as state-of-the-art among AR-based methods or as competitive with state-of-the-art.
  3. [Section 4.3, Tables 1b and 2] The ablation design conflates the two proposed components and reveals an inconsistency. The Baseline in Table 2 differs from Ours by both hierarchical tokenization and DPO, while the w/o DPO entry in Table 1b has hierarchical tokenization but no DPO. Comparing these shows that hierarchical tokenization alone can degrade full-resolution PSNR: on RealSR at 512, the Table 2 Baseline (no hierarchical, no DPO) reaches PSNR 26.11, while the Table 1b w/o DPO (hierarchical, no DPO) reaches 25.72. Thus the improvement attributed to the tokenizer in Table 2 is actually due to the combination with DPO, and the conclusion that hierarchical tokenization is beneficial 'even for the target of fixed-scale super resolution' is not supported. Please provide a complete 2x2 ablation (with/without hierarchical tokenization crossed with with/without DPO) or explicitly describe the comparison as a joint ablation.
  4. [Section 3, Hierarchical RQ-VAE paragraph] The training procedure for the hierarchical RQ-VAE is described inconsistently. The text first says 'we finetune the vocabulary and the decoder of the RQ-VAE' and that the finetuning incorporates a scale-specific decoder with standard reconstruction losses for each scale, but two sentences later it says 'We then keep the decoder frozen and update the vocabulary using the gradient of the l2 distance...'. This leaves unclear whether the scale-specific decoder is trained by the reconstruction losses and at what point it is frozen. Because the multi-scale decoding capability is a central claim, please specify the exact sequence of training stages and which parameters are updated in each stage.
minor comments (5)
  1. [Section 4.1] The stated compression factor f=0.25 is inconsistent with the reported 1024 conditioning tokens for a 512x512 input and the final resolution rho_L=32: with f=0.25, the 512x512 input would produce a 128x128 feature map and 16,384 tokens. Please reconcile these numbers (likely f=0.0625 or a different final resolution).
  2. [Algorithm 1] The notation in Algorithm 1, particularly the conditions involving 'max_k rho_k <= s_i rho_L' and 'i < max_k rho_k <= s_{i-1} rho_L', is very hard to parse. Please rewrite the stopping condition and the loop indices with explicit definitions.
  3. [Section 4.1] The text says degradation uses 'a small probability p<=0.25' for pure bilinear downsampling, but the exact value of p is never given. Please state the value used in the experiments.
  4. [Table 3] The table caption says 'Red and blue colors represent best and second-best results,' but the table as typeset contains no red or blue highlighting. Either add the colors or remove the sentence.
  5. [Figure 1 caption] The caption contains a typo: 'Hierarchical VQV AE Tokenization' should presumably be 'Hierarchical RQ-VAE Tokenization'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: intermediate-scale decoding is a trained capability and DPO is a training objective benchmarked externally.

full rationale

The paper's central claims are empirical and architectural rather than derivational. The hierarchical RQ-VAE is explicitly finetuned with multi-scale reconstruction losses and evaluated on held-out ImageNet-512 and standard SR benchmarks, so the ability to decode intermediate scales is a trained capability, not a consequence of how the evaluation is defined. The DPO regularizer in Eq. (4) is a training objective whose effect is measured by ablations on RealSR and DRealSR; although the manuscript under-specifies exactly how p(zLR) is scored (which conditioning prefix is used for the logits), that is a reproducibility gap, not circularity, because the reported gains are not implied by the definition of the loss. Background self-citations (e.g., [23], [24]) are not load-bearing for any of the main claims. The comparisons against VARSR, diffusion-based, and GAN-based baselines are external, and no parameter is fitted to the test set and then renamed as a prediction. No circular step was found.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper rests on standard machine learning machinery plus pre-trained checkpoints and an ad hoc preference loss. No physical entities or constants are postulated.

free parameters (3)
  • beta in DPO loss = not reported
    Eq. 4 scales the log-ratio; magnitude was not stated. Its value controls the strength of the preference term.
  • probability p of pure bilinear downsampling = <=0.25
    Used during training with RealESRGAN degradations; exact value not reported.
  • DPO vs cross-entropy loss weight = 1:1 equal weights
    Section 3 states equal weights; no sensitivity analysis is given.
assumptions (4)
  • domain assumption RealESRGAN degradation approximates real-world degradations
    Training and the synthetic test set both use the RealESRGAN pipeline, so the metric on the synthetic benchmark partly reflects the training distribution. Section 4.2.
  • domain assumption Pre-trained Switti and VAR d-16 checkpoints provide useful, transferable priors
    The RQ-VAE is initialized from Switti and the transformer from VAR d-16; the method assumes these checkpoints are publicly available and suitable. Section 4.1.
  • domain assumption Partial residual sequences decode to valid intermediate-scale images after vocabulary and decoder finetuning
    Core to multi-scale decoding; supported by Table 1a but not formally guaranteed. Section 3 and Table 1a.
  • ad hoc to paper A preference loss without a reference model is a valid training objective
    The DPO term drops the reference model, so it is not the standard DPO; it is a ranking loss adapted for this task. Eq. 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-scale Image Super Resolution with a Single Auto-Regressive Model." pith.science (2026). https://pith.science/paper/XKJDAI2C

@misc{pith2026250604990,
  author       = {Pith},
  title        = {Pith review of: Multi-scale Image Super Resolution with a Single Auto-Regressive Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XKJDAI2C}},
  note         = {Machine review of arXiv:2506.04990}
}
read the original abstract

In this paper we tackle Image Super Resolution (ISR), using recent advances in Visual Auto-Regressive (VAR) modeling. VAR iteratively estimates the residual in latent space between gradually increasing image scales, a process referred to as next-scale prediction. Thus, the strong priors learned during pre-training align well with the downstream task (ISR). To our knowledge, only VARSR has exploited this synergy so far, showing promising results. However, due to the limitations of existing residual quantizers, VARSR works only at a fixed resolution, i.e. it fails to map intermediate outputs to the corresponding image scales. Additionally, it relies on a 1B transformer architecture (VAR-d24), and leverages a large-scale private dataset to achieve state-of-the-art results. We address these limitations through two novel components: a) a Hierarchical Image Tokenization approach with a multi-scale image tokenizer that progressively represents images at different scales while simultaneously enforcing token overlap across scales, and b) a Direct Preference Optimization (DPO) regularization term that, relying solely on the LR and HR tokenizations, encourages the transformer to produce the latter over the former. To the best of our knowledge, this is the first time a quantizer is trained to force semantically consistent residuals at different scales, and the first time that preference-based optimization is used to train a VAR. Using these two components, our model can denoise the LR image and super-resolve at half and full target upscale factors in a single forward pass. Additionally, we achieve \textit{state-of-the-art results on ISR}, while using a small model (300M params vs ~1B params of VARSR), and without using external training data.

Figures

Figures reproduced from arXiv: 2506.04990 by the authors.

Figure 1
Figure 1. Top: we represent the quantization of an input image (up to L = 6 scales for clarity of visualization). We observe that reconstructing the image using all 6 residuals leads to perfect reconstruction. However, mapping the residuals up to the first three scales does not result in the desired 2×-downsampled version of the input image, i.e., the initial scales do not convey scale￾wise semantic information of the encoded… view at source ↗
Figure 2
Figure 2. Multi-scale SR evaluation of (a) VARSR, (b) Baseline, and (c) Our proposed approach. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results: (Top) without and (Bottom) with our proposed DPO-based regulariza￾tion. We can see the role of the regularization term in sharpening results. Dataset Metrics 128 256 512 VARSR Baseline Ours VARSR Baseline Ours VARSR Baseline Ours DIV2K-Val PSNR↑ 18.58 17.61 21.95 20.19 19.55 24.04 23.91 24.69 24.20 SSIM↑ 0.4086 0.3811 0.5949 0.5010 0.4868 0.6656 0.5980 0.6470 0.6226 LPIPS↓ 0.5501 0.6530 0.2092 0… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative results. (a) Input LR (upsampled to target resolution); (b) Ground truth; (c) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Multi-scale SR evaluation of (a) VARSR, (b) Baseline, and (c) Our proposed approach. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results. (a) Input LR (upsampled to target resolution); (b) Ground truth; (c) [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results. (a) Input LR (upsampled to target resolution); (b) Ground truth; (c) [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 38 canonical work pages

  1. [1]

    NTIRE 2017 challenge on single image super-resolution: Dataset and study

    Eirikur Agustsson and Radu Timofte. NTIRE 2017 challenge on single image super-resolution: Dataset and study. InIEEE Conference on Computer Vision and Pattern Recognition - Workshops, 2017

    work page 2017

  2. [2]

    Toward real-world single image super-resolution: A new benchmark and a new model

    Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. InIEEE International Conference on Computer Vision, 2019

    work page 2019

  3. [3]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. InIEEE Conference on Computer Vision and Pattern Recognition, 2009

    work page 2009

  4. [4]

    Learning a deep convolutional network for image super-resolution

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Learning a deep convolutional network for image super-resolution. InEuropean Conference on Computer Vision, 2014

    work page 2014

  5. [5]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. InIEEE Conference on Computer Vision and Pattern Recognition, 2021

    work page 2021

  6. [6]

    Fluid: Scaling autoregressive text-to-image generative models with continuous tokens

    Lijie Fan, Tianhong Li, Siyang Qin, Yuanzhen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. InInternational Conference on Learning Representations, 2025

    work page 2025

  7. [7]

    R. Gray. Vector quantization.IEEE ASSP Magazine, 1984

    work page 1984

  8. [8]

    Image processing using multi-code GAN prior

    Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code GAN prior. InIEEE Conference on Computer Vision and Pattern Recognition, 2020

    work page 2020

Show all 44 references
  1. [9]

    DIV8K: DIVerse 8K resolution image dataset

    Shuhang Gu, Andreas Lugmayr, Martin Danelljan, Manuel Fritsche, Julien Lamour, and Radu Timofte. DIV8K: DIVerse 8K resolution image dataset. InIEEE International Conference on Computer Vision - Workshops, 2019

    2019

  2. [10]

    Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. InIEEE Conference on Computer Vision and Pattern Recognition, 2025

    2025

  3. [11]

    Improving autoregressive visual generation with cluster-oriented token prediction, 2025

    Teng Hu, Jiangning Zhang, Ran Yi, Jieyu Weng, Yabiao Wang, Xianfang Zeng, Zhucun Xue, and Lizhuang Ma. Improving autoregressive visual generation with cluster-oriented token prediction, 2025

    2025

  4. [12]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InIEEE Conference on Computer Vision and Pattern Recognition, 2019

    2019

  5. [13]

    Accurate image super-resolution using very deep convolutional networks

    Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. InIEEE Conference on Computer Vision and Pattern Recognition, 2016

    2016

  6. [14]

    The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale.International Journal on Computer Vision, 2020

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection a...

    2020

  7. [15]

    Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi

    Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew P. Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. InIEEE Conferenc...

    2017

  8. [16]

    Autoregressive image generation using residual quantization

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. InIEEE Conference on Computer Vision and Pattern Recognition, 2022

    2022

  9. [17]

    Autoregressive image generation without vector quantization

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. InNeural Information Processing Systems, 2024

    2024

  10. [18]

    Swinir: Image restoration using swin transformer.arXiv preprint arXiv:2108.10257, 2021

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer.arXiv preprint arXiv:2108.10257, 2021

    2021 arXiv

  11. [19]

    Blind image superresolution: A survey and beyond

    Anran Liu, Yihao Liu, Jinjin Gu, Yu Qiao, and Chao Dong. Blind image superresolution: A survey and beyond. InarXiv preprint arXiv:2107.03055, 2021

    2021 arXiv

  12. [20]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2019. 10

    2019

  13. [21]

    STAR: Scale-wise text-to-image generation via auto-regressive representations.arXiv preprint arXiv:2406.10797, 2024

    Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. STAR: Scale-wise text-to-image generation via auto-regressive representations.arXiv preprint arXiv:2406.10797, 2024

    2024 arXiv

  14. [22]

    CoDi: Conditional diffusion distillation for higher-fidelity and faster image generation.IEEE Conference on Computer Vision and Pattern Recognition, 2024

    Kangfu Mei, Mauricio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. CoDi: Conditional diffusion distillation for higher-fidelity and faster image generation.IEEE Conference on Computer Vision and Pattern Recognition, 2024

    2024

  15. [23]

    Edge-sd-sr: Low latency and parameter efficient on-device super-resolution with stable diffusion via bidirectional conditioning

    Mehdi Noroozi, Isma Hadji, Victor Escorcia, Anestis Zaganidis, Brais Martinez, and Georgios Tzimiropou- los. Edge-sd-sr: Low latency and parameter efficient on-device super-resolution with stable diffusion via bidirectional conditioning. InIEEE Conference on Computer Vision an...

    2025

  16. [24]

    You only need one step: Fast super-resolution with stable diffusion via scale distillation

    Mehdi Noroozi, Isma Hadji, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. You only need one step: Fast super-resolution with stable diffusion via scale distillation. InEuropean Conference on Computer Vision, 2024

    2024

  17. [25]

    Visual autoregressive modeling for image super-resolution

    Yunpeng Qu, Kun Yuan, Jinhua Hao, Kai Zhao, Qizhi Xie, Ming Sun, and Chao Zhou. Visual autoregressive modeling for image super-resolution. InInternational Conference on Machine Learning, 2025

    2025

  18. [26]

    Language models are unsupervised multitask learners.OpenAI blog, 1(8):9, 2019

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners.OpenAI blog, 1(8):9, 2019

    2019

  19. [27]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. InNeural Information Processing Systems, 2023

    2023

  20. [28]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  21. [29]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 2024

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 2024

    2024

  22. [30]

    Autore- gressive model beats diffusion: Llama for scalable image generation.arXiv preprint arXiv:2406.06525, 2024

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autore- gressive model beats diffusion: Llama for scalable image generation.arXiv preprint arXiv:2406.06525, 2024

    2024 arXiv

  23. [31]

    HART: efficient visual generation with hybrid autoregressive transformer

    Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. HART: efficient visual generation with hybrid autoregressive transformer. International Conference on Learning Representations, 2025

    2025

  24. [32]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. InNeural Information Processing Systems, 2024

    2024

  25. [33]

    NTIRE 2017 challenge on single image super-resolution: Methods and results

    Radu Timofte, Eirikur Agustsson, Luc Van Gool, MingHsuan Yang, and Lei Zhang. NTIRE 2017 challenge on single image super-resolution: Methods and results. InIEEE Conference on Computer Vision and Pattern Recognition - Workshops, 2017

    2017

  26. [34]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. Neural Information Processing Systems, 2017

    2017

  27. [35]

    Switti: Designing scale-wise transformers for text-to-image synthesis

    Anton V oronov, Denis Kuznedelev, Mikhail Khoroshikh, Valentin Khrulkov, and Dmitry Baranchuk. Switti: Designing scale-wise transformers for text-to-image synthesis. InIEEE Conference on Computer Vision and Pattern Recognition, 2025

    2025

  28. [36]

    Chan, and Chen Change Loy

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution.International Journal on Computer Vision, 2024

    2024

  29. [37]

    Real-ESRGAN: Training real-world blind super- resolution with pure synthetic data

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training real-world blind super- resolution with pure synthetic data. InIEEE International Conference on Computer Vision - Workshops, 2021

    2021

  30. [38]

    Recovering realistic texture in image super- resolution by deep spatial feature transform

    Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super- resolution by deep spatial feature transform. InIEEE Conference on Computer Vision and Pattern Recognition, 2018

    2018

  31. [39]

    Deep networks for image super-resolution with sparse prior

    Zhaowen Wang, Ding Liu, Jianchao Yang, Wei Han, and Thomas Huang. Deep networks for image super-resolution with sparse prior. InIEEE International Conference on Computer Vision, 2015. 11

    2015

  32. [40]

    Component divide-and-conquer for real-world image super-resolution

    Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. InEuropean Conference on Computer Vision, 2020

    2020

  33. [41]

    Resshift: Efficient diffusion model for image super-resolution by residual shifting

    Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. InNeural Information Processing Systems, 2023

    2023

  34. [42]

    Designing a practical degradation model for deep blind image super-resolution

    Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. InIEEE International Conference on Computer Vision, 2021

    2021

  35. [43]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. InIEEE International Conference on Computer Vision, 2023

    2023

  36. [44]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InIEEE Conference on Computer Vision and Pattern Recognition, 2018. Appendix In Fig. 5 we present more examples of the limitati...

    2018

Pith tools

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