Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

One-step DiT super-resolution works when trajectory mismatch and high-frequency spectral leakage are regularized together.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Asymmetric adversarial distillation plus frequency distribution matching lets DiT models perform one-step Real-ISR without the grid-like periodic artifacts that plague prior one-step DiT distillations.

T0 review reviewed 2026-07-15 challenge →

load-bearing objection Practical one-step DiT Real-ISR recipe that kills the grid artifacts and posts clean SOTA perceptual numbers on the usual benchmarks. the 3 major comments →

arxiv 2603.06275 v2 pith:YFSC7VZY submitted 2026-03-06 cs.CV

Spectral and Trajectory Regularization for Diffusion Transformer Super-Resolution

Jingkai Wang , Yixin Tang , Jue Gong , Jiatong Li , Shu Li , Libo Liu , Jianliang Lan , Yutong Liu
show 1 more author
Yulun Zhang
This is my paper
classification cs.CV
keywords diffusion transformerreal-world image super-resolutionone-step distillationtrajectory regularizationfrequency distribution matchingspectral leakageadversarial distillation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Diffusion transformers can generate high-quality real-world super-resolution, but their multi-step sampling is too slow for practical use and one-step distillation of them produces severe grid-like artifacts. The paper argues that the root causes are a trajectory mismatch between the pretrained noise-to-image path and the new low-resolution-to-high-resolution path, plus high-frequency spectral leakage inside DiT patches. StrSR solves both with a single training recipe: an asymmetric discriminator based on a pretrained CLIP-ConvNeXt that keeps training stable, and a frequency distribution matching loss that aligns amplitude and phase spectra. On standard Real-ISR benchmarks the resulting one-step models recover photo-realistic textures that earlier DiT and UNet distillations miss, while remaining as fast as other one-step methods. A sympathetic reader cares because this shows large DiT backbones can be unlocked for efficient image restoration once the right spectral and trajectory constraints are added.

Core claim

The authors claim that one-step real-world super-resolution with diffusion transformers becomes practical and artifact-free once an asymmetric discriminative distillation (CLIP-ConvNeXt RaGAN plus approximated R1) bridges the multi-step to one-step trajectory gap and a frequency distribution matching loss (sliced Wasserstein on amplitude and phase) suppresses the high-frequency spectral leakage that produces DiT-specific grid artifacts.

What carries the argument

Frequency distribution loss (FDL) together with asymmetric RaGAN distillation: FDL minimizes sliced Wasserstein distance between predicted and target feature spectra in both amplitude and phase; the asymmetric discriminator supplies stable adversarial pressure that a same-architecture DiT discriminator cannot.

Load-bearing premise

A single large Euler step starting from a VAE-encoded low-resolution latent can be forced onto the clean high-resolution distribution by the proposed losses alone, without progressive multi-stage distillation or an explicit trajectory corrector.

What would settle it

Train the identical dual-encoder DiT pipeline with the same LoRA rank and data but without FDL and without the CLIP-ConvNeXt RaGAN; if the resulting images still show the same dense periodic grid artifacts and the same LPIPS/DISTS gap versus multi-step teachers on RealSR and DIV2K-val, the central claim that these two regularizers are what close the trajectory and spectral gaps collapses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes StrSR, a one-step adversarial distillation framework for real-world image super-resolution (Real-ISR) based on diffusion transformers (DiT). It diagnoses two DiT-specific obstacles to one-step generation: trajectory mismatch between the pretrained noise-to-data path and the LR-to-HR path (Fig. 2, §3.1), and grid-like periodic artifacts attributed to high-frequency spectral leakage from DiT patchification (Fig. 3, §3.4). The remedies are (i) an asymmetric discriminative distillation setup that freezes a CLIP-ConvNeXt backbone as a PatchGAN-style discriminator, trained with relativistic average GAN loss plus approximated R1 regularization (§3.3, Eqs. 7–9), and (ii) a frequency distribution loss (FDL) that matches sliced Wasserstein distances on DFT amplitude and phase of intermediate features (§3.4, Eq. 10). A dual-encoder (VAE + frozen VLM) supplies spatial and semantic conditioning. Experiments on DIV2K-val, RealSR and RealLQ250 with two DiT backbones (Z-Image-Turbo, FLUX.2-klein) report best or second-best LPIPS/DISTS among one-step methods and competitive no-reference scores, supported by ablations of FDL, RaGAN and VLM components.

Significance. If the empirical claims hold, StrSR is a practically useful contribution: it shows that large pretrained DiTs can be distilled into competitive one-step Real-ISR models without progressive multi-stage distillation, while specifically mitigating the grid artifacts that appear more severely in DiT than in UNet architectures. The combination of an asymmetric CLIP-ConvNeXt discriminator with spectral sliced-Wasserstein matching is a concrete, reusable design pattern. Strengths include honest reporting of the expected PSNR/SSIM trade-off for generative SR, consistent gains across two distinct DiT backbones and three datasets, visual evidence of artifact reduction (Figs. 6, 10–11), and a promised code/model release. The work is therefore of clear interest to the Real-ISR and efficient diffusion communities.

major comments (3)
  1. §3.1 and Fig. 2 treat the VAE-encoded LR latent as an intermediate state zt1 on a rectified-flow trajectory, then take a single Euler step. The paper never states how t1 is chosen or whether it is fixed across degradation severities. Because the claimed trajectory gap widens under severe degradation, the sensitivity of final LPIPS/DISTS (and residual artifacts) to t1 should be reported; otherwise it is unclear whether the asymmetric RaGAN + FDL losses alone close the gap or whether a carefully tuned t1 is an unstated free parameter.
  2. Tables 1–3 and the ablation in Table 4 report point estimates without error bars, multiple random seeds, or statistical significance tests. Given that several one-step baselines are already close on no-reference metrics (e.g., TSD-SR, HYPIR), the SOTA ranking on MANIQA/MUSIQ/QAlign is not yet robustly established. At minimum, standard deviations over 3–5 seeds for the key perceptual metrics on one dataset would strengthen the central claim.
  3. §3.4 attributes the grid artifacts to high-frequency spectral leakage and introduces FDL (Eq. 10) as the remedy. The supporting evidence is qualitative (Figs. 6, 10). A quantitative spectral analysis—e.g., power-spectrum distance or measured period of residual grid energy before/after FDL—would make the causal link between spectral leakage and the observed artifacts more rigorous and would better justify the specific choice of sliced Wasserstein on amplitude and phase.
minor comments (6)
  1. §3.1 heading “Preliminarys” is misspelled; correct to “Preliminaries”.
  2. Eq. (8) writes D(·) while the surrounding text uses C(·) for discriminator logits; unify the notation.
  3. Fig. 4 caption and the main text both use ℰVAE / 𝒟VAE and EVAE / DVAE interchangeably; pick one symbol set.
  4. In Tables 1–3 the abbreviations MAN., MUS., QAli., Qual. are defined only in the caption; a short legend or expanded column headers would improve readability.
  5. The two-stage schedule (40 k steps without FDL, then 20 k with λ3 = 0.002) is stated in §4.1 but never ablated; a short note on whether joint training from the start is stable would be useful.
  6. Related-work coverage of concurrent one-step DiT SR methods (TSD-SR, FluxSR, DiT4SR) is adequate, but a brief discussion of why progressive distillation was deliberately avoided (beyond the SeedVR2 citation) would help readers place the design choice.

Circularity Check

0 steps flagged

No significant circularity: empirical one-step DiT Real-ISR method whose claims rest on external held-out metrics and ablations, not definitional or self-referential reductions.

full rationale

StrSR is a standard empirical computer-vision paper. The trajectory-mismatch diagnosis (Sec. 3.1, Fig. 2) and DiT spectral-leakage diagnosis (Fig. 3, Sec. 3.4) motivate two concrete regularizers: an asymmetric CLIP-ConvNeXt RaGAN + approximated R1 discriminator (Eqs. 7–9) and frequency distribution matching via sliced Wasserstein on amplitude/phase (Eq. 10). The generator objective (Eq. 11) simply combines these losses with ordinary L1 + LPIPS. All quantitative claims (Tables 1–3) and visual claims (Figs. 6–11) are evaluated on held-out DIV2K-val, RealSR and RealLQ250 against independently published baselines; ablations (Table 4, Figs. 10–11) isolate each component. No equation is definitionally equivalent to its target, no parameter is fitted on a subset and then “predicted” on a near-identical quantity, and no uniqueness theorem or load-bearing premise is imported solely via self-citation. Self-citations that appear are ordinary background references to prior SR or distillation work and are not used to force the central result. Consequently the derivation chain contains no circular steps.

Axiom & Free-Parameter Ledger

6 free parameters · 3 axioms · 2 invented entities

The central empirical claim rests on standard rectified-flow and GAN machinery plus a handful of hand-chosen loss weights and architectural choices; no new physical entities are postulated. Free parameters are the usual training hyper-parameters of modern generative models.

free parameters (6)
  • λ1 (LPIPS weight) = 3
    Set to 3 by hand; controls perception-distortion trade-off in generator objective (Eq. 11).
  • λ2 (RaGAN weight) = 0.1
    Set to 0.1; balances adversarial signal against reconstruction losses.
  • λ3 (FDL weight) = 0 / 0.002
    0 in stage 1, 0.002 in stage 2; controls strength of spectral regularization.
  • λ_R1 = 10
    Set to 10; strength of approximated R1 regularizer on discriminator.
  • LoRA rank = 256
    Chosen as 256 for generator fine-tuning; capacity hyper-parameter.
  • training steps / stages = 40k + 20k
    40 k steps without FDL then 20 k with FDL; schedule chosen empirically.
axioms (3)
  • domain assumption Rectified-flow linear trajectory zt = t z1 + (1-t) z0 yields a well-defined vector field that can be distilled into a one-step Euler map from LR latent to HR latent.
    Invoked throughout §3.1 and Fig. 2; standard in recent flow-matching SR papers but not proved for severe real-world degradations.
  • domain assumption A pretrained CLIP-ConvNeXt backbone remains a reliable texture-sensitive discriminator after only lightweight convolutional heads are trained.
    Core of the asymmetric design in §3.3; relies on transferability of CLIP features to Real-ISR artifacts.
  • ad hoc to paper Sliced Wasserstein distance on DFT amplitude and phase of intermediate features is a sufficient spectral regularizer to suppress DiT patch-induced periodic artifacts.
    Introduced as FDL in §3.4; empirically motivated but not derived from first principles of DiT spectral leakage.
invented entities (2)
  • StrSR asymmetric discriminative distillation architecture no independent evidence
    purpose: Bridge multi-step to one-step trajectory gap while avoiding DiT discriminator collapse.
    New combination of frozen CLIP-ConvNeXt + RaGAN + approx R1; no independent theoretical existence proof outside the paper’s experiments.
  • Frequency distribution loss (FDL) for DiT SR no independent evidence
    purpose: Penalize high-frequency spectral leakage that produces grid artifacts.
    Adaptation of prior FDL work to the DiT one-step setting; effectiveness shown only inside this paper’s ablations.

reviewed 2026-07-15 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Spectral and Trajectory Regularization for Diffusion Transformer Super-Resolution." pith.science (2026). https://pith.science/paper/YFSC7VZY

@misc{pith2026260306275,
  author       = {Pith},
  title        = {Pith review of: Spectral and Trajectory Regularization for Diffusion Transformer Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YFSC7VZY}},
  note         = {Machine review of arXiv:2603.06275}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Diffusion transformer (DiT) architectures show great potential for real-world image super-resolution (Real-ISR). However, their computationally expensive iterative sampling necessitates one-step distillation. Existing one-step distillation methods struggle with Real-ISR on DiT. They suffer from fundamental trajectory mismatch and generate severe grid-like periodic artifacts. To tackle these challenges, we propose StrSR, a novel one-step adversarial distillation framework featuring spectral and trajectory regularization. Specifically, we propose an asymmetric discriminative distillation architecture to bridge the trajectory gap. Additionally, we design a frequency distribution matching strategy to effectively suppress DiT-specific periodic artifacts caused by high-frequency spectral leakage. Extensive experiments demonstrate that StrSR achieves state-of-the-art performance in Real-ISR, across both quantitative metrics and visual perception. The code and models will be released at https://github.com/jkwang28/StrSR .

Figures

Figures reproduced from arXiv: 2603.06275 by Jianliang Lan, Jiatong Li, Jingkai Wang, Jue Gong, Libo Liu, Shu Li, Yixin Tang, Yulun Zhang, Yutong Liu.

Figure 1
Figure 1. Figure 1: For a self-captured cat image with dense fur, TSD-SR [16] and FluxSR [28] suffer from heavy grid-like artifacts, whereas our StrSR restores realistic details. However, generating high-quality images with diffusion models usually requires dozens or hundreds of steps, which is computationally expensive. To tackle this issue, many works [16, 59, 75, 77] distill pretrained text-to-image (T2I) models for image-… view at source ↗
Figure 2
Figure 2. Figure 2: Two gaps exist between the multi-step T2I model and the one-step SR model. First, a direct linear trajectory from the intermediate step t = t1 to t = 0 cannot reach the real HR data distribution. Second, the LR data distribution mismatches the pretrained T2I model’s distribution at t1. Points in the figure represent distributions. where v(xt, t) is the time-dependent vector field. By solving this ODE from … view at source ↗
Figure 3
Figure 3. Figure 3: One-step T2I inference along a linear trajectory: UNet vs. DiT. While both models produce artifacts, the DiT-based model exhibits more severe grid-like artifacts. target in a single Euler step: zˆ0 = zt1 − Gθ(zt1 , t1, c). When image degradation is mild, the mapping distance between the pretrained noise trajectory Tnoise and the new restoration trajectory TLR is relatively small. In such cases, fine-tuning… view at source ↗
Figure 4
Figure 4. Figure 4: Training pipeline of StrSR. The LR image xLR is input into the dual-encoder EVAE and EVLM and following DiT Gθ to gain the predicted HR image xpred. DiT is fine-tuned by LoRA. The generator and discriminator are trained alternately. In parallel, xLR is processed by a pretrained VAE encoder, EVAE, to obtain the continuous latent space representation zVAE, i.e., zLR: z_\text {LR} = \mathcal {E}_\text {VAE}(x… view at source ↗
Figure 5
Figure 5. Figure 5: Discriminator comparison. ConvNeXt’s large-kernel convolutions preserve multi￾scale spatial information. Conversely, DiT’s patchification heavily compresses intra-patch details (edges and textures), reducing its ability to detect patch-level periodic artifacts. understanding, and superior texture perception. Most importantly, the strong local inductive bias of ConvNeXt makes it highly sensitive to high-fre… view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of spectral decomposition with and without FDL. Without FDL, distinct periodic artifacts are visible in the left image. After applying FDL, the spectral leakage is effectively mitigated. Type Methods PSNR↑ SSIM↑ LPIPS↓ DISTS↓ NIQE↓ MAN.↑ MUS.↑ QAli.↑ Qual.↑ Multi-step Diffusion SUPIR [76] 22.22 0.5748 0.3806 0.1660 3.4372 0.5793 63.4906 4.3862 0.6959 DiT4SR [17] 20.76 0.5451 0.3995 0.1670 3.0084… view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparison for image SR (×4) in DIV2K-val dataset. Qualitative Results Visual comparisons in Figs. 7, 8, and 9 show that StrSR yields superior visual results. We select StrSR based on the FLUX.2 [klein] 4B as a pretrained model for these visual comparisons. For the synthetic DIV2K-val dataset, StrSR correctly recognizes high-level semantics to generate accurate texture categories. Furthermore, it pr… view at source ↗
Figure 8
Figure 8. Figure 8: Visual comparison for image SR (×4) in RealSR dataset. 4.3 Ablation Study Frequency Distribution Matching. We extract features from the final layer of the DiT to visualize the impact of the FDL. As shown in [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual comparison for image SR (×4) in RealLQ250 dataset. The first column is the bicubic algorithm. Please zoom in for a better view. w/o FDL PCA Norm Cluster Heatmap w/ FDL PCA Norm Cluster Heatmap [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visual comparison for extracted feature map. We extract the features of the generator’s last layer for demonstration, including four widely-used visualization methods. Our FDL loss effectively alleviates artifacts in latent level. VLM Dual Encoder. Specifically, we evaluate the model under three conditions: without any text embedding (w/o emb), with a naive text encoder (w/o VLM), and with our VLM-encoded… view at source ↗
Figure 11
Figure 11. Figure 11: (Left) Comparison of image textures with and without frequency distribu￾tion matching. FDL guides the model to generate more photo-realistic fine details. (Right) Visual comparison of the ablation results for the GAN and VLM components. StrSR accurately captures both semantic information and texture details. Methods MAN. MUS. QAli. w/o GAN 0.4720 54.086 4.1629 w/o Ra 0.4930 61.820 4.4570 w/o emb 0.5043 58… view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of inference times for image super-resolution (×4) across various methods. The vertical axis is displayed on a logarithmic scale. Our StrSR, although utilizing a significantly larger back￾bone, achieves comparable inference speeds. 4.4 Running Time We further evaluate the inference speed for 1,024×1,024 input and compare it with various SR methods on an NVIDIA RTX A6000 GPU, as shown in [PITH_… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. FoA-SR: Faithful or Aesthetic? Profile-Aware Preference Optimization for Real-World Image Super-Resolution

    cs.CV 2026-06 unverdicted novelty 6.0

    FoA-SR trains separate Faithful and Aesthetic LoRA adapters by ranking a shared stochastic candidate pool with profile-specific rewards after initial supervised Flux2SR training.

Reference graph

Works this paper leans on

83 extracted references · 14 linked inside Pith · cited by 1 Pith paper

  1. [1]

    arXiv preprint arXiv:2403.11176 (2024)

    Agnolucci, L., Galteri, L., Bertini, M.: Quality-aware image-text alignment for opinion-unaware image quality assessment. arXiv preprint arXiv:2403.11176 (2024)

    Pith/arXiv arXiv 2024

  2. [2]

    In: CVPR (2017)

    Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image super-resolution: Dataset and study. In: CVPR (2017)

    2017

  3. [3]

    In: NeurIPS (2024)

    Ai, Y., Zhou, X., Huang, H., Han, X., Chen, Z., You, Q., Yang, H.: DreamClear: High-capacity real-world image restoration with privacy-safe dataset curation. In: NeurIPS (2024)

    2024

  4. [4]

    Black Forest Labs: Flux.https://github.com/black-forest-labs/flux

  5. [5]

    https://huggingface.co/black- forest- labs/FLUX.2-klein-4B

    Black Forest Labs: Flux.2-klein-4b. https://huggingface.co/black- forest- labs/FLUX.2-klein-4B

  6. [6]

    In: CVPR (2018)

    Blau, Y., Michaeli, T.: The perception-distortion tradeoff. In: CVPR (2018)

    2018

  7. [7]

    In: ICCV (2019)

    Cai, J., Zeng, H., Yong, H., Cao, Z., Zhang, L.: Toward real-world single image super-resolution: A new benchmark and a new model. In: ICCV (2019)

    2019

  8. [8]

    In: AAAI (2025)

    Chadebec, C., Tasar, O., Benaroche, E., Aubin, B.: Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. In: AAAI (2025)

    2025

  9. [9]

    [Online]

    Chen, C., Mo, J.: IQA-PyTorch: Pytorch toolbox for image quality assessment. [Online]. Available:https://github.com/chaofengc/IQA-PyTorch

  10. [10]

    In: CVPR (2025)

    Chen, D.Y., Bandyopadhyay, H., Zou, K., Song, Y.Z.: Nitrofusion: High-fidelity single-step diffusion through dynamic adversarial training. In: CVPR (2025)

    2025

  11. [11]

    IEEE TPAMI (2025)

    Chen, X., Wang, X., Zhang, W., Kong, X., Qiao, Y., Zhou, J., Dong, C.: HAT: Hybrid attention transformer for image restoration. IEEE TPAMI (2025)

    2025

  12. [12]

    In: ICCV (2023)

    Chen,Z.,Zhang,Y.,Gu,J.,Kong,L.,Yang,X.,Yu,F.:Dualaggregationtransformer for image super-resolution. In: ICCV (2023)

    2023

  13. [13]

    In: CVPR (2023)

    Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: CVPR (2023)

    2023

  14. [14]

    IEEE TPAMI (2020)

    Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. IEEE TPAMI (2020)

    2020

  15. [15]

    IEEE TPAMI (2016)

    Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolu- tional networks. IEEE TPAMI (2016)

    2016

  16. [16]

    In: CVPR (2025)

    Dong, L., Fan, Q., Guo, Y., Wang, Z., Zhang, Q., Chen, J., Luo, Y., Zou, C.: TSD-SR: One-step diffusion with target score distillation for real-world image super-resolution. In: CVPR (2025)

    2025

  17. [17]

    In: ICCV (2025)

    Duan, Z.P., Zhang, J., Jin, X., Zhang, Z., Xiong, Z., Zou, D., Ren, J.S., Guo, C., Li, C.: DiT4SR: Taming diffusion transformer for real-world image super-resolution. In: ICCV (2025)

    2025

  18. [18]

    In: ICML (2024)

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: ICML (2024)

    2024

  19. [19]

    In: NeurIPS (2014)

    Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)

    2014

  20. [20]

    arXiv preprint arXiv:2010.14701 (2020)

    Henighan, T., Kaplan, J., Katz, M., Chen, M., Hesse, C., Jackson, J., Jun, H., Brown, T.B., Dhariwal, P., Gray, S., et al.: Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701 (2020)

    Pith/arXiv arXiv 2010

  21. [21]

    In: NeurIPS (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)

    2020

  22. [22]

    In: ICLR (2022) 16 J

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022) 16 J. Wang et al

    2022

  23. [23]

    arXiv preprint arXiv:1807.00734 (2018)

    Jolicoeur-Martineau, A.: The relativistic discriminator: A key element missing from standard gan. arXiv preprint arXiv:1807.00734 (2018)

    Pith/arXiv arXiv 2018

  24. [24]

    In: ECCV (2024)

    Kang, M., Zhang, R., Barnes, C., Paris, S., Kwak, S., Park, J., Shechtman, E., Zhu, J.Y., Park, T.: Distilling diffusion models into conditional gans. In: ECCV (2024)

    2024

  25. [25]

    In: ICCV (2021)

    Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: MUSIQ: Multi-scale Image Quality Transformer . In: ICCV (2021)

    2021

  26. [26]

    In: ICLR (2014)

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014)

    2014

  27. [27]

    In: CVPR (2017)

    Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image super-resolution using a generative adversarial network. In: CVPR (2017)

    2017

  28. [28]

    In: ICML (2025)

    Li, J., Cao, J., Guo, Y., Li, W., Zhang, Y.: One diffusion step to real-world super- resolution via flow trajectory distillation. In: ICML (2025)

    2025

  29. [29]

    In: NeurIPS (2025)

    Li, J., Cao, J., Zou, Z., Su, X., Yuan, X., Zhang, Y., Guo, Y., Yang, X.: Unleashing the power of one-step diffusion based image super-resolution via a large-scale diffusion discriminator. In: NeurIPS (2025)

    2025

  30. [30]

    In: CVPR (2023)

    Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., et al.: LSDIR: A large scale dataset for image restoration. In: CVPR (2023)

    2023

  31. [31]

    In: ICCVW (2021)

    Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: ICCVW (2021)

    2021

  32. [32]

    arXiv preprint arXiv:2501.08316 (2025)

    Lin, S., Xia, X., Ren, Y., Yang, C., Xiao, X., Jiang, L.: Diffusion adversarial post-training for one-step video generation. arXiv preprint arXiv:2501.08316 (2025)

    arXiv 2025

  33. [33]

    In: ECCV (2024)

    Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., Ouyang, W., Qiao, Y., Dong, C.: DiffBIR: Towards blind image restoration with generative diffusion prior. In: ECCV (2024)

    2024

  34. [34]

    In: SIGGRAPH Asia (2025)

    Lin, X., Yu, F., Hu, J., You, Z., Shi, W., Ren, J.S., Gu, J., Dong, C.: Harnessing diffusion-yielded score priors for image restoration. In: SIGGRAPH Asia (2025)

    2025

  35. [35]

    In: ICLR (2023)

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: ICLR (2023)

    2023

  36. [36]

    In: NeurIPS (2025)

    Liu, S., Ma, J., Sun, L., Kong, X., Zhang, L.: InstructRestore: Region-customized image restoration with human instructions. In: NeurIPS (2025)

    2025

  37. [37]

    arXiv preprint arXiv:2209.03003 (2022)

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003 (2022)

    Pith/arXiv arXiv 2022

  38. [38]

    In: CVPR (2022)

    Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: CVPR (2022)

    2022

  39. [39]

    arXiv preprint arXiv:1711.05101 (2017)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    Pith/arXiv arXiv 2017

  40. [40]

    In: ICLR (2019)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)

    2019

  41. [41]

    arXiv preprint arXiv:2410.11081 (2024)

    Lu, C., Song, Y.: Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081 (2024)

    Pith/arXiv arXiv 2024

  42. [42]

    arXiv preprint arXiv:2310.04378 (2023)

    Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023)

    Pith/arXiv arXiv 2023

  43. [43]

    In: NeurIPS (2023)

    Luo, W., Hu, T., Zhang, S., Sun, J., Li, Z., Zhang, Z.: Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. In: NeurIPS (2023)

    2023

  44. [44]

    arXiv preprint arXiv:2403.12931 (2024)

    Luo, Y., Chen, X., Qu, X., Hu, T., Tang, J.: You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. arXiv preprint arXiv:2403.12931 (2024)

    Pith/arXiv arXiv 2024

  45. [45]

    In: CVPR (2025) Spectral and Trajectory Regularization for DiT SR 17

    Nguyen, T.H., Tran, A.: Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In: CVPR (2025) Spectral and Trajectory Regularization for DiT SR 17

    2025

  46. [46]

    In: CVPR (2024)

    Ni, Z., Wu, J., Wang, Z., Yang, W., Wang, H., Ma, L.: Misalignment-robust frequency distribution loss for image transformation. In: CVPR (2024)

    2024

  47. [47]

    In: ICCV (2023)

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023)

    2023

  48. [48]

    arXiv preprint arXiv:2307.01952 (2023)

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)

    Pith/arXiv arXiv 2023

  49. [49]

    arXiv preprint arXiv:2505.09388 (2025)

    Qwen Team: Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025)

    Pith/arXiv arXiv 2025

  50. [50]

    In: ICML (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)

    2021

  51. [51]

    In: CVPR (2022)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)

    2022

  52. [52]

    In: NeurIPS (2017)

    Roth, K., Lucchi, A., Nowozin, S., Hofmann, T.: Stabilizing training of generative adversarial networks through regularization. In: NeurIPS (2017)

    2017

  53. [53]

    arXiv preprint arXiv:2202.00512 (2022)

    Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512 (2022)

    Pith/arXiv arXiv 2022

  54. [54]

    In: SIGGRAPH Asia (2024)

    Sauer, A., Boesel, F., Dockhorn, T., Blattmann, A., Esser, P., Rombach, R.: Fast high-resolution image synthesis with latent adversarial diffusion distillation. In: SIGGRAPH Asia (2024)

    2024

  55. [55]

    In: ECCV (2024)

    Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation. In: ECCV (2024)

    2024

  56. [56]

    In: ICLR (2021)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)

    2021

  57. [57]

    arXiv preprint arXiv:2310.14189 (2023)

    Song, Y., Dhariwal, P.: Improved techniques for training consistency models. arXiv preprint arXiv:2310.14189 (2023)

    Pith/arXiv arXiv 2023

  58. [58]

    In: ICML (2023)

    Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. In: ICML (2023)

    2023

  59. [59]

    In: CVPR (2025)

    Sun, L., Wu, R., Ma, Z., Liu, S., Yi, Q., Zhang, L.: Pixel-level and semantic-level adjustable super-resolution: A dual-lora approach. In: CVPR (2025)

    2025

  60. [60]

    In: ICLR (2026)

    Wang, J., Lin, S., Lin, Z., Ren, Y., Wei, M., Yue, Z., Zhou, S., Chen, H., Zhao, Y., Yang, C., et al.: SeedVR2: One-step video restoration via diffusion adversarial post-training. In: ICLR (2026)

    2026

  61. [61]

    In: ICCVW (2021)

    Wang, X., Xie, L., Dong, C., Shan, Y.: Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. In: ICCVW (2021)

    2021

  62. [62]

    In: ECCVW (2018)

    Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: ESRGAN: Enhanced super-resolution generative adversarial networks. In: ECCVW (2018)

    2018

  63. [63]

    In: CVPR (2024)

    Wang, Y., Yang, W., Chen, X., Wang, Y., Guo, L., Chau, L.P., Liu, Z., Qiao, Y., Kot, A.C., Wen, B.: SinSR: diffusion-based image super-resolution in a single step. In: CVPR (2024)

    2024

  64. [64]

    In: NeurIPS (2023)

    Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: NeurIPS (2023)

    2023

  65. [65]

    arXiv preprint arXiv:2508.02324 (2025)

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., ming Yin, S., Bai, S., Xu, X., Chen, Y., Chen, Y., Tang, Z., Zhang, Z., Wang, Z., Yang, A., Yu, B., Cheng, C., Liu, D., Li, D., Zhang, H., Meng, H., Wei, H., Ni, J., Chen, K., Cao, K., Peng, L., Qu, L., Wu, M., Wang, P., Yu, S., Wen, T., Feng, W., Xu, X., Wang, Y., Zhang, Y., Zhu, Y., Wu, Y., Cai, Y., L...

    Pith/arXiv arXiv 2025

  66. [66]

    In: ICML (2024) 18 J

    Wu, H., Zhang, Z., Zhang, W., Chen, C., Li, C., Liao, L., Wang, A., Zhang, E., Sun, W., Yan, Q., Min, X., Zhai, G., Lin, W.: Q-Align: Teaching lmms for visual scoring via discrete text-defined levels. In: ICML (2024) 18 J. Wang et al

    2024

  67. [67]

    In: NeurIPS (2024)

    Wu, R., Sun, L., Ma, Z., Zhang, L.: One-step effective diffusion network for real- world image super-resolution. In: NeurIPS (2024)

    2024

  68. [68]

    In: CVPR (2024)

    Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., Zhang, L.: SeeSR: Towards semantics- aware real-world image super-resolution. In: CVPR (2024)

    2024

  69. [69]

    In: ICCV (2023)

    Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: DiffIR: Efficient diffusion model for image restoration. In: ICCV (2023)

    2023

  70. [70]

    In: CVPR (2024)

    Xu, Y., Zhao, Y., Xiao, Z., Hou, T.: Ufogen: You forward once large scale text-to- image generation via diffusion gans. In: CVPR (2024)

    2024

  71. [71]

    In: CVPRW (2022)

    Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y., Cao, M., Wang, J., Yang, Y.: MANIQA: Multi-dimension attention network for no-reference image quality assessment. In: CVPRW (2022)

    2022

  72. [72]

    In: ECCV (2024)

    Yang, T., Wu, R., Ren, P., Xie, X., Zhang, L.: Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In: ECCV (2024)

    2024

  73. [73]

    In: NeurIPS (2024)

    Yin, T., Gharbi, M., Park, T., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T.: Improved distribution matching distillation for fast image synthesis. In: NeurIPS (2024)

    2024

  74. [74]

    In: CVPR (2024)

    Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. In: CVPR (2024)

    2024

  75. [75]

    In: ICCV (2025)

    You, W., Zhang, M., Zhang, L., Zhou, X., Shi, K., Gu, S.: Consistency trajectory matching for one-step generative super-resolution. In: ICCV (2025)

    2025

  76. [76]

    In: CVPR (2024)

    Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. In: CVPR (2024)

    2024

  77. [77]

    In: CVPR (2025)

    Yue, Z., Liao, K., Loy, C.C.: Arbitrary-steps image super-resolution via diffusion inversion. In: CVPR (2025)

    2025

  78. [78]

    arXiv preprint arXiv:2511.22699 (2025)

    Z-Image Team: Z-Image: An efficient image generation foundation model with single-stream diffusion transformer. arXiv preprint arXiv:2511.22699 (2025)

    Pith/arXiv arXiv 2025

  79. [79]

    In: CVPR (2025)

    Zhang,J.,Huang,Q.,Liu,J.,Guo,X.,Huang,D.:Diffusion-4k:Ultra-high-resolution image synthesis with latent diffusion models. In: CVPR (2025)

    2025

  80. [80]

    In: ICCV (2021)

    Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: ICCV (2021)

    2021

Showing first 80 references.

This paper was first reviewed by grok-4.5 on July 15, 2026.