Pith. sign in

REVIEW 6 major objections 5 minor 82 references

Latent Guidance in Diffusion Models for Perceptual Evaluations

T0 review · 6 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 The paper claims that steering a pretrained latent diffusion model's sampling with perceptual-quality features produces intermediate features aligned with human quality judgments, and that reading out multi-scale, multi-timestep features…

desk verdict Solid empirical contribution in NR-IQA, but the manifold-guidance theory is unsupported and should be reframed. read the letter →

arxiv 2506.00327 v1 pith:N62NOXJF submitted 2025-05-31 cs.CV cs.AI

Shreshth Saini , Ru-Ling Liao , Yan Ye , Alan C. Bovik This is my paper
classification cs.CVcs.AI
keywords no-referenceimagequalityassessmentlatentdiffusionmodelsperceptualmanifoldguidancehyperfeaturesconsistencyon-manifoldsamplingDDIM
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

The paper sets out to show that a pretrained latent diffusion model, used without any fine-tuning, already contains the information needed to judge image quality the way humans do. It proposes guiding the model's reverse sampling steps with weak perceptual-quality features, then reading out multi-scale, multi-timestep feature maps from the denoising U-Net. A small regression network turns these 'diffusion hyperfeatures' into a quality score. On ten authentic, synthetic, and AI-generated image-quality datasets, this pipeline reports correlations with human ratings that match or exceed those of prior specialist models, and it transfers across datasets better than they do.

What carries the argument

The load-bearing object is Perceptual Manifold Guidance (PMG), an extra update inserted into DDIM sampling. At each reverse step the algorithm forms Tweedie's estimate $z_{0|t}$, applies a data-consistency gradient $G_1$ and a perceptual-feature gradient $G_2$ computed from a perceptual model $\psi_p$, and only then performs the DDIM step; Theorem 1 is the assertion that these two gradients, if they lie in the tangent space of the latent manifold, keep the whole trajectory on the manifold and push the estimate into a perceptually consistent sub-region. The second mechanism is diffusion hyperfeatures: the union of selected denoising U-Net features across layers and timesteps, which a two-layer regression head maps to a quality score.

What would settle it

Run LGDM on the same images with the real Stable Diffusion VAE and with a near-lossless, invertible autoencoder, measuring both the decoded distance of the guided estimate $D(z''_{0|t})$ from the clean input and the resulting SRCC. If the accuracy gain from perceptual guidance over no guidance ($\zeta_2 = 0$) persists unchanged when the reconstruction error is driven to zero, the on-manifold mechanism is not what produces the improvement; if the gain disappears or shrinks sharply, the lossy VAE was doing the work that Theorem 1 assigns to the manifold structure.

Watch

Extended reading notes

Core claim

The central claim is that latent diffusion models implicitly organize their latent space into perceptually consistent regions, and that steering the sampling trajectory toward those regions makes the U-Net's internal features align with human quality judgments. Concretely, the Perceptual Manifold Guidance update acts on the estimated clean latent with two $\ell^2$ losses, one for content consistency with the input and one for consistency of perceptual features, and the paper argues (Theorem 1, under a perfect-autoencoder and tangent-space assumption) that all updates stay on noisy latent manifolds, with the guided estimate landing in a perceptually consistent locality. Aggregating features across U-Net layers and diffusion timesteps then yields hyperfeatures whose correlation with human opinion is reported as PLCC/SRCC 0.940/0.908 on LIVEC, 0.812/0.705 on FLIVE, 0.921/0.883 on TID2013, and 0.961/0.958 on KADID, with the same recipe also leading on cross-dataset transfers and on AI-generated-content benchmarks.

Load-bearing premise

The proof that guided samples stay in a perceptually consistent region assumes the autoencoder is perfect and the guidance gradients lie in the latent manifold's tangent space, but the real Stable Diffusion VAE is lossy; if those conditions fail, the theoretical guarantee has no bite and the reported gains rest on empirical evidence alone.

Editorial extensions

If this is right

  • A frozen latent diffusion model can serve as a general-purpose no-reference quality feature extractor, removing the need to train or fine-tune a quality backbone per dataset.
  • Because the features span scales and timesteps, the method captures both fine texture artifacts and larger structural distortions, which supports its transfer across authentic, synthetic, and AI-generated content.
  • The perceptual-guidance recipe transfers to different plug-in feature extractors, boosting those that already correlate with human judgment and degrading performance when the plug-in is weak.
  • Cross-dataset transfer is reported to be stronger than for fine-tuned comparators, including when training on a small dataset and testing on a much larger one.
  • Inference cost is the main practical trade-off: ten sampling steps give near-best accuracy at about 21 seconds per image, while a one-step variant is much faster but less accurate.

Reading between the lines

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

  • If the on-manifold guarantee is the real driver of the gains, then using a near-lossless autoencoder should preserve most of LGDM's accuracy; a lossy VAE could be introducing exactly the off-manifold drift the theory assumes away, so this is a testable consequence the paper does not run.
  • The two-term guidance recipe is generic: any low-level vision task whose error can be written as a differentiable content loss plus a perceptual-feature loss could reuse PMG and the same hyperfeature readout, for example denoising, super-resolution, or quality-aware generation.
  • The optimal perceptual-guidance weight $\zeta_2$ likely tracks the plug-in metric's own agreement with human opinion, which would let practitioners predict the best weight for a new $\psi_p$ from its baseline correlation instead of tuning on each dataset.
  • The observed dominance of one U-Net layer suggests a cheaper operational variant that extracts only the most predictive layers and timesteps; the paper identifies the trade-off but does not implement this reduced-cost version.
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

6 major / 5 minor

Summary. The paper proposes LGDM, a no-reference image quality assessment (NR-IQA) method built on a frozen pretrained latent diffusion model. The method introduces Perceptual Manifold Guidance (PMG), which modifies the DDIM sampling trajectory in latent space by adding a data-consistency term and a perceptual-consistency term, and then extracts multi-scale, multi-timestep U-Net features (diffusion hyperfeatures) that are regressed to quality scores. Experiments on ten authentic, synthetic, and AI-generated IQA datasets report state-of-the-art PLCC/SRCC results, with ablations on the choice of perceptual metric, number of timesteps, guidance weight, Stable Diffusion version, and cross-dataset generalization.

Significance. If the mechanism claimed by the paper were established, the contribution would be significant: it would show that a zero-shot pretrained latent diffusion model, without fine-tuning, can provide perceptual features that outperform specialized NR-IQA models across diverse distortions, and it would give a theoretical account of on-manifold perceptual guidance. The empirical study is a clear strength: ten datasets, cross-dataset evaluations, multiple ablations, and a frozen backbone are reported, and the final quality scores are regressed against human MOS, so the headline empirical claim is not circular. However, the theoretical support for the central on-manifold claim is not sound in its present form, and the best-performing variant uses the same network for perceptual guidance and for hyperfeature extraction, which complicates the interpretation of the perceptual-guidance mechanism.

major comments (6)
  1. [Appendix B.5, Eq. (B-35)] The proof of Proposition 2 states ∂E/∂x0 · ∂D/∂z0 = I, but with E: R^D → R^k and D: R^k → R^D, the product is k×k and cannot equal the D×D identity. The correct chain-rule identity for a perfect autoencoder is ∂D/∂z0 · ∂E/∂x0 = I_D. As written, the identity is dimensionally inconsistent, so the proof that D(∇G) lies in T_{x0}M is invalid.
  2. [Theorem 1 and Appendix B.7, Eq. (B-44)–(B-46)] The theorem assumes that the gradients ∇G1 and ∇G2 reside in the tangent space T_{z0|t}Z, which is exactly the property the theorem is supposed to establish for the updated latent. Without an independent argument that these gradients are tangent, the conclusion that all update terms remain on the latent manifold is circular.
  3. [Theorem 1 and Appendix E] The on-manifold guarantee assumes a perfect encoder-decoder pair, but the paper itself acknowledges in Appendix E that real VAEs 'introduce notable reconstruction errors.' Since the experiments use the Stable Diffusion VAE, the theoretical guarantee does not apply to the actual system used in the reported results. The empirical results may still stand, but the claim that PMG provably keeps samples in a perceptually consistent manifold locality is not supported.
  4. [Lemma 2, Appendix B.6] The concentration argument requires an optimal score function, ϵθ = ϵ_t, which is an idealized condition that does not hold for the trained Stable Diffusion model. The proof also does not address how the noise term in Eq. (B-44) interacts with the non-tangent component of the gradient when the score is not optimal, so the concentration of z_{t-1} on Z_{t-1} is only established under an assumption known to be violated.
  5. [Eq. (B-46) and Section 3.1] The set M_perceptual is introduced only textually as a sub-manifold of the content-consistent manifold; no formal definition, metric, or proof is given that z''_{0|t} lies in M_perceptual rather than merely in the tangent space. Since the paper's novelty is perceptual consistency, this is a load-bearing gap rather than a presentation issue.
  6. [Appendix C.2 and Table 1] For the best variant LGDM-ψSDv1.5, the perceptual guidance target ψp is computed by the same Stable Diffusion network whose intermediate features form the hyperfeatures H. Although the features are computed before the guidance loop and are stated to be distinct, the guidance loss then aligns the sampled latent with features from the same network used for prediction. This makes it difficult to attribute the large gain over LGDM-ψφ to human-perceptual alignment rather than to self-consistency of the network's own features. An additional ablation using a different network for ψp, with matched capacity, would clarify the mechanism.
minor comments (5)
  1. [Abstract and Section 3.2] The phrase 'multi-time features' is imprecise; 'multi-timestep' would be clearer.
  2. [Tables 1 and 4] Dataset names are inconsistent: 'KONIQ' in Table 1 and 'KonIQ' in Table 4 both refer to KonIQ-10k.
  3. [References, Li et al. 2024a] The reference title contains a typo: 'AIGIQA-20k' should be 'AGIQA-20k'.
  4. [Appendix C.3] The text says the Stable Diffusion VAE is a VQ-VAE, but Stable Diffusion v1.5 uses a KL-regularized autoencoder; this should be corrected.
  5. [Appendix C.1] The subjective evaluation is said to follow ITU-T P.910, which is a video quality recommendation; for still images, ITU-R BT.500 or a comparable image-specific protocol would be the standard citation.

Circularity Check

2 steps flagged · score 4.0 of 10

Empirical NR-IQA results are externally benchmarked, but Theorem 1's on-manifold guarantee assumes the tangent-space property it claims to prove.

  1. other [Theorem 1 (Section 3.1) and proof in Appendix B.7, Eqs. B-44 to B-46]
    "Given Assumption 1, given a perfect encoder E, decoder D, and an efficient score function sθ(zt, t), let the gradient ∇z0|t G1(D(z0|t), y) and ∇z0|t G2(ψp(D(z0|t)), ψp(y)) reside on the tangent space Tz0|t Z of the latent manifold Z. Throughout the diffusion process, all update terms zt remain on noisy latent manifolds Zt, with z′′0|t lying in a perceptually consistent manifold locality."

    The theorem's conclusion is its own premise. The proof in Eq. B-44 updates z0|t by subtracting ζ1∇G1 + ζ2∇G2, so z′0|T1 stays in Z only because both gradients were assumed to lie in Tz0|T1 Z; no argument establishes that the PMG gradients are tangent. The 'perceptually consistent manifold locality' conclusion is likewise the G2 loss by construction: z′′0|t is defined as a gradient step on G2(ψp(D(·)), ψp(y)), so asserting that it lies in a perceptually consistent region restates the update rule rather than deriving a property. Thus the advertised guarantee that PMG keeps sampling on the data manifold reduces to an assumption, not a proof.

  2. self definitional [Appendix C.2 (ψSDv1.5 calculation) and Algorithm 1, lines 9-12]
    "In the case of ψSDv1.5, a special case where we leverage the SDv1.5 model to pre-compute the perceptual features. Crucially, this calculation is done separately and beforehand, and the resulting features are distinct from the H collected later. One can think of these features as having a similar structure as H, and calculated in a similar way, but without line 9 in Algorithm 1."

    For the best-performing variant, the perceptual guidance target ψp(y) is computed by the same SDv1.5 U-Net that later supplies the hyperfeatures H. Algorithm 1 line 9 aligns D(ˆz0|t) to ψp(x) in the feature space of this same network, and line 12 regresses H to MOS. The 'perceptual consistency' enforced by Eq. 10 is therefore self-referential: the sampler is guided to match features generated by the very model whose features are then evaluated. This does not force the MOS correlation, but it means the perceptual-consistency mechanism for LGDM-ψSDv1.5 is self-alignment rather than an independent perceptual ground truth.

full rationale

The reported NR-IQA scores are not circular: the final quality prediction qp = gϕ(H) is fit to human MOS on standard benchmarks, so the SOTA numbers in Tables 1-3 are externally grounded. The circularity lies in the paper's theoretical mechanism. Theorem 1 assumes that both PMG gradients lie in the tangent space Tz0|t Z and then concludes that all update terms remain on the latent manifold; that is the central property PMG is claimed to guarantee, so the proof is tantamount to assuming its own conclusion. The same theorem also assumes a perfect encoder-decoder pair and an optimal score function, while Appendix E concedes that 'real VAEs introduce notable reconstruction errors.' Additionally, Proposition 2's proof states ∂E/∂x0 · ∂D/∂z0 = I, which is dimensionally impossible for E: R^D → R^k and D: R^k → R^D with k < D; this is a correctness flaw rather than a circularity, but it reinforces that the on-manifold mechanism is unsupported. The ψSDv1.5 variant adds a self-referential element because the perceptual target and the evaluated hyperfeatures come from the same network, though this does not determine the MOS regression outcome. Because the central empirical claim is independently benchmarked, the score is 4 rather than 6 or higher.

Assumptions & free parameters 3 free parameters · 6 assumptions · 1 invented entities

The central empirical claim rests mainly on the fitted regression head and on the pretrained Stable Diffusion backbone. The theoretical on-manifold guarantee rests on a perfect-autoencoder assumption and on an assumed tangent-space condition, both acknowledged as unrealistic or unproven. The perceptual guidance weights and timestep range are tuned on validation data. M_perceptual is an invented theoretical object with no independent evidence.

free parameters (3)
  • Perceptual guidance weights ζ1, ζ2 = ζ1=1, ζ2=0.2
    Chosen empirically; ablation in Appendix D.3 shows 0.2 is optimal on LIVEC and FLIVE. The method's performance depends on these values.
  • Number of DDIM steps T and timestep range = T=10, t in (0,100]
    Selected from ablations in Table 5 and Fig. 7; more steps improve SRCC but increase cost, and larger timestep ranges degrade performance.
  • Regression head g_phi weights = Trained on 70% train split of each target dataset
    The quality prediction q_p = g_phi(H) is a supervised regressor trained on human MOS. The final NR-IQA numbers depend on this fitted model, so the method is not zero-shot end to end.
assumptions (6)
  • domain assumption Assumption 1: Strong Manifold Hypothesis; data lies on a low-dimensional locally linear subspace manifold.
    Stated in Section 2.1 and used throughout Proposition 1, Proposition 2, and Theorem 1.
  • ad hoc to paper Perfect autoencoder: x = D(E(x)) with exact inverse Jacobians.
    Used in Proposition 2 and Theorem 1 (Appendix B.5-B.7). Appendix E admits real VAEs have notable reconstruction errors, so this assumption is known to be false in practice.
  • ad hoc to paper Guidance gradients ∇G1 and ∇G2 reside in the tangent space Tz0|t Z of the latent manifold.
    This is a hypothesis of Theorem 1 (Appendix B.7) and is not proven; the theorem's conclusion depends on it directly.
  • domain assumption Conditional independence of y and ψp(y) given zt for later diffusion timesteps.
    Used in Appendix B.3 to split the conditional score into two likelihood terms; justified only heuristically.
  • domain assumption Pretrained latent diffusion models implicitly learn perceptually consistent manifold regions.
    Core hypothesis stated in Sections 1 and 3. Empirically supported by ablations but not proven independently.
  • standard math Standard diffusion and score-matching theory: DDIM updates, Tweedie's formula, VP-SDE reverse process.
    Background theory in Section 2.2 and Appendix A-B, cited from prior work (Song et al., Ho et al., Efron).
invented entities (1)
  • M_perceptual, a perceptually consistent sub-manifold within the content-consistent manifold
    purpose: The region of the latent manifold that PMG claims to steer samples toward; used to state Theorem 1 and justify Equation 10.
    Defined by assertion in the Theorem 1 proof (Appendix B.7, Eq. B-46) as a subset of M_content. There is no falsifiable handle outside the paper that identifies or measures this sub-manifold independently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent Guidance in Diffusion Models for Perceptual Evaluations." pith.science (2026). https://pith.science/paper/N62NOXJF

@misc{pith2026250600327,
  author       = {Pith},
  title        = {Pith review of: Latent Guidance in Diffusion Models for Perceptual Evaluations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N62NOXJF}},
  note         = {Machine review of arXiv:2506.00327}
}
read the original abstract

Despite recent advancements in latent diffusion models that generate high-dimensional image data and perform various downstream tasks, there has been little exploration into perceptual consistency within these models on the task of No-Reference Image Quality Assessment (NR-IQA). In this paper, we hypothesize that latent diffusion models implicitly exhibit perceptually consistent local regions within the data manifold. We leverage this insight to guide on-manifold sampling using perceptual features and input measurements. Specifically, we propose Perceptual Manifold Guidance (PMG), an algorithm that utilizes pretrained latent diffusion models and perceptual quality features to obtain perceptually consistent multi-scale and multi-timestep feature maps from the denoising U-Net. We empirically demonstrate that these hyperfeatures exhibit high correlation with human perception in IQA tasks. Our method can be applied to any existing pretrained latent diffusion model and is straightforward to integrate. To the best of our knowledge, this paper is the first work on guiding diffusion model with perceptual features for NR-IQA. Extensive experiments on IQA datasets show that our method, LGDM, achieves state-of-the-art performance, underscoring the superior generalization capabilities of diffusion models for NR-IQA tasks.

Figures

Figures reproduced from arXiv: 2506.00327 by the authors.

Figure 1
Figure 1. An overview of our proposed approach: (a) shows [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Exemplar Authentic and Synthetic distorted im [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Illustrates aggregation of multi-scale and multi [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Effect of ζ2 on the hyperfeatures. Figure on the left and right, shows the PLCC and SRCC values respec￾tively for FLIVE and LIVEC datasets. Hyperfeatures were collected for single time step. Impact of Weights of Perceptual Guidance Term (ζ). We conducted experiments to…
Figure 6
Figure 6. Figure 6: Contribution of individual layer of SDv1.5 to [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: We report the behaviour of LGDM as we change the range of timesteps in the sampling process. As we move [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

82 extracted references · 59 canonical work pages

  1. [1]

    Arniqa: Learning distortion manifold for image quality assessment

    Agnolucci, L., Galteri, L., Bertini, M., and Del Bimbo, A. Arniqa: Learning distortion manifold for image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp.\ 189--198, 2024

    2024

  2. [2]

    ediffiqa: Towards efficient face image quality assessment based on denoising diffusion probabilistic models

    Babnik, Z ., Peer, P., and S truc, V. ediffiqa: Towards efficient face image quality assessment based on denoising diffusion probabilistic models. IEEE Transactions on Biometrics, Behavior, and Identity Science, 2024

    2024

  3. [3]

    Label-efficient semantic segmentation with diffusion models

    Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., and Babenko, A. Label-efficient semantic segmentation with diffusion models. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=SlxSY2UZQT

    work page 2022

  4. [4]

    The manifold hypothesis for gradient-based explanations

    Bordt, S., Upadhyay, U., Akata, Z., and von Luxburg, U. The manifold hypothesis for gradient-based explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 3697--3702, 2023

    work page 2023

  5. [5]

    Deep neural networks for no-reference and full-reference image quality assessment

    Bosse, S., Maniry, D., Müller, K.-R., Wiegand, T., and Samek, W. Deep neural networks for no-reference and full-reference image quality assessment. IEEE Transactions on Image Processing, 27 0 (1): 0 206--219, 2018. doi:10.1109/TIP.2017.2760518

    work page arXiv 2018

  6. [6]

    Chung, H., Sim, B., Ryu, D., and Ye, J. C. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 35: 0 25683--25696, 2022

    work page 2022

  7. [7]

    T., Klasky, M

    Chung, H., Kim, J., Mccann, M. T., Klasky, M. L., and Ye, J. C. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=OnD9zGAGT0k

    2023

  8. [8]

    Chung, H., Lee, S., and Ye, J. C. Decomposed diffusion sampler for accelerating large-scale inverse problems. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=DsEhqQtfAG

    work page 2024

Show all 82 references
  1. [9]

    Genziqa: Generalized image quality assessment using prompt-guided latent diffusion models

    De, D., Mitra, S., and Soundararajan, R. Genziqa: Generalized image quality assessment using prompt-guided latent diffusion models. arXiv preprint arXiv:2406.04654, 2024

    2024

  2. [10]

    Tweedie’s formula and selection bias

    Efron, B. Tweedie’s formula and selection bias. Journal of the American Statistical Association, 106 0 (496): 0 1602--1614, 2011

    2011

  3. [11]

    Perceptual quality assessment of smartphone photography

    Fang, Y., Zhu, H., Zeng, Y., Ma, K., and Wang, Z. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3677--3686, 2020

    2020

  4. [12]

    Generative diffusion prior for unified image restoration and enhancement

    Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., and Dai, B. Generative diffusion prior for unified image restoration and enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9935--9946, 2023

    2023

  5. [13]

    Semantically structured image compression via irregular group-based decoupling

    Feng, R., Gao, Y., Jin, X., Feng, R., and Chen, Z. Semantically structured image compression via irregular group-based decoupling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 17237--17247, 2023

    2023

  6. [14]

    Dp-iqa: Utilizing diffusion prior for blind image quality assessment in the wild

    Fu, H., Wang, Y., Yang, W., and Wen, B. Dp-iqa: Utilizing diffusion prior for blind image quality assessment in the wild. arXiv preprint arXiv:2405.19996, 2024

    2024

  7. [15]

    Local manifold learning for no-reference image quality assessment

    Gao, T., Pan, W., Zhang, Y., Zhao, S., Zhang, S., Zheng, X., Li, K., Cao, L., and Ji, R. Local manifold learning for no-reference image quality assessment. arXiv preprint arXiv:2406.19247, 2024

    2024

  8. [16]

    and Bovik, A

    Ghadiyaram, D. and Bovik, A. C. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25 0 (1): 0 372--387, 2015

    2015

  9. [17]

    Foundation models boost low-level perceptual similarity metrics

    Ghildyal, A., Barman, N., and Zadtootaghaj, S. Foundation models boost low-level perceptual similarity metrics. arXiv preprint arXiv:2409.07650, 2024

    2024 arXiv

  10. [18]

    A., Dadsetan, S., and Kitani, K

    Golestaneh, S. A., Dadsetan, S., and Kitani, K. M. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp.\ 1220--1230, 2022

    2022

  11. [19]

    No-reference high-dynamic-range image quality assessment based on tensor decomposition and manifold learning

    Guan, F., Jiang, G., Song, Y., Yu, M., Peng, Z., and Chen, F. No-reference high-dynamic-range image quality assessment based on tensor decomposition and manifold learning. Applied Optics, 57 0 (4): 0 839--848, 2018

    2018

  12. [20]

    Visual importance and distortion guided deep image quality assessment framework

    Guan, J., Yi, S., Zeng, X., Cham, W.-K., and Wang, X. Visual importance and distortion guided deep image quality assessment framework. IEEE Transactions on Multimedia, 19 0 (11): 0 2505--2520, 2017

    2017

  13. [21]

    Z., Salakhutdinov, R., and Ermon, S

    He, Y., Murata, N., Lai, C.-H., Takida, Y., Uesaka, T., Kim, D., Liao, W.-H., Mitsufuji, Y., Kolter, J. Z., Salakhutdinov, R., and Ermon, S. Manifold preserving guided diffusion. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview....

    2024

  14. [22]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  15. [23]

    Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment

    Hosu, V., Lin, H., Sziranyi, T., and Saupe, D. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29: 0 4041--4056, 2020

    2020

  16. [24]

    Global structure-aware diffusion process for low-light image enhancement

    Hou, J., Zhu, Z., Hou, J., Liu, H., Zeng, H., and Yuan, H. Global structure-aware diffusion process for low-light image enhancement. Advances in Neural Information Processing Systems, 36, 2024

    2024

  17. [25]

    Subjective video quality assessment methods for multimedia applications

    ITU-T RECOMMENDATION, P. Subjective video quality assessment methods for multimedia applications. 1999

    1999

  18. [26]

    Unified no-reference quality assessment of singly and multiply distorted stereoscopic images

    Jiang, Q., Shao, F., Gao, W., Chen, Z., Jiang, G., and Ho, Y.-S. Unified no-reference quality assessment of singly and multiply distorted stereoscopic images. IEEE Transactions on Image Processing, 28 0 (4): 0 1866--1881, 2018

    2018

  19. [27]

    Denoising diffusion restoration models

    Kawar, B., Elad, M., Ermon, S., and Song, J. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35: 0 23593--23606, 2022

    2022

  20. [28]

    Musiq: Multi-scale image quality transformer

    Ke, J., Wang, Q., Wang, Y., Milanfar, P., and Yang, F. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 5148--5157, 2021

    2021

  21. [29]

    Larson, E. C. and Chandler, D. M. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 19 0 (1): 0 011006--011006, 2010

    2010

  22. [30]

    Agiqa-3k: An open database for ai-generated image quality assessment

    Li, C., Zhang, Z., Wu, H., Sun, W., Min, X., Liu, X., Zhai, G., and Lin, W. Agiqa-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023

    2023

  23. [31]

    Aigiqa-20k: A large database for ai-generated image quality assessment

    Li, C., Kou, T., Gao, Y., Cao, Y., Sun, W., Zhang, Z., Zhou, Y., Zhang, Z., Zhang, W., Wu, H., et al. Aigiqa-20k: A large database for ai-generated image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 6327--6336, 2024 a

    2024

  24. [32]

    Feature denoising diffusion model for blind image quality assessment

    Li, X., Zheng, J., Hu, R., Zhang, Y., Li, K., Shen, Y., Zheng, X., Liu, Y., Zhang, S., Dai, P., et al. Feature denoising diffusion model for blind image quality assessment. arXiv preprint arXiv:2401.11949, 2024 b

    2024 arXiv

  25. [33]

    Kadid-10k: A large-scale artificially distorted iqa database

    Lin, H., Hosu, V., and Saupe, D. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pp.\ 1--3. IEEE, 2019

    2019

  26. [34]

    Liqa: Lifelong blind image quality assessment

    Liu, J., Zhou, W., Li, X., Xu, J., and Chen, Z. Liqa: Lifelong blind image quality assessment. IEEE Transactions on Multimedia, 25: 0 5358--5373, 2022

    2022

  27. [35]

    Learned image compression with mixed transformer-cnn architectures

    Liu, J., Sun, H., and Katto, J. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 14388--14397, 2023

    2023

  28. [36]

    H., Holynski, A., and Darrell, T

    Luo, G., Dunlap, L., Park, D. H., Holynski, A., and Darrell, T. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024

    2024

  29. [37]

    C., Birkbeck, N., Wang, Y., Adsumilli, B., and Bovik, A

    Madhusudana, P. C., Birkbeck, N., Wang, Y., Adsumilli, B., and Bovik, A. C. Image quality assessment using contrastive learning. IEEE Transactions on Image Processing, 31: 0 4149--4161, 2022

    2022

  30. [38]

    completely blind

    Mittal, A., Soundararajan, R., and Bovik, A. C. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20 0 (3): 0 209--212, 2012

    2012

  31. [39]

    Moorthy, A. K. and Bovik, A. C. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE transactions on Image Processing, 20 0 (12): 0 3350--3364, 2011

    2011

  32. [40]

    Understanding the latent space of diffusion models through the lens of riemannian geometry

    Park, Y.-H., Kwon, M., Choi, J., Jo, J., and Uh, Y. Understanding the latent space of diffusion models through the lens of riemannian geometry. Advances in Neural Information Processing Systems, 36: 0 24129--24142, 2023

    2023

  33. [41]

    Color image database tid2013: Peculiarities and preliminary results

    Ponomarenko, N., Ieremeiev, O., Lukin, V., Egiazarian, K., Jin, L., Astola, J., Vozel, B., Chehdi, K., Carli, M., Battisti, F., et al. Color image database tid2013: Peculiarities and preliminary results. In European workshop on visual information processing (EUVIP), pp.\ 106--...

    2013

  34. [42]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  35. [43]

    Solving linear inverse problems provably via posterior sampling with latent diffusion models

    Rout, L., Raoof, N., Daras, G., Caramanis, C., Dimakis, A., and Shakkottai, S. Solving linear inverse problems provably via posterior sampling with latent diffusion models. Advances in Neural Information Processing Systems, 36, 2024

    2024

  36. [44]

    A., Bovik, A

    Saad, M. A., Bovik, A. C., and Charrier, C. Blind image quality assessment: A natural scene statistics approach in the dct domain. IEEE transactions on Image Processing, 21 0 (8): 0 3339--3352, 2012

    2012

  37. [45]

    Saha, A., Mishra, S., and Bovik, A. C. Re-iqa: Unsupervised learning for image quality assessment in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 5846--5855, 2023

    2023

  38. [46]

    Saini, S., Saha, A., and Bovik, A. C. Hidro-vqa: High dynamic range oracle for video quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp.\ 469--479, 2024

    2024

  39. [47]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35: 0...

    2022

  40. [48]

    The riemannian geometry of deep generative models

    Shao, H., Kumar, A., and Thomas Fletcher, P. The riemannian geometry of deep generative models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp.\ 315--323, 2018

    2018

  41. [49]

    R., Sabir, M

    Sheikh, H. R., Sabir, M. F., and Bovik, A. C. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image processing, 15 0 (11): 0 3440--3451, 2006

    2006

  42. [50]

    Blind image quality assessment based on geometric order learning

    Shin, N.-H., Lee, S.-H., and Kim, C.-S. Blind image quality assessment based on geometric order learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12799--12808, 2024

    2024

  43. [51]

    M., Zhang, Z., Hu, X., Qu, Q., and Shen, L

    Song, B., Kwon, S. M., Zhang, Z., Hu, X., Qu, Q., and Shen, L. Solving inverse problems with latent diffusion models via hard data consistency. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=j8hdRqOUhN

    2024

  44. [52]

    Denoising diffusion implicit models

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021 a . URL https://openreview.net/forum?id=St1giarCHLP

    2021

  45. [53]

    Loss-guided diffusion models for plug-and-play controllable generation

    Song, J., Zhang, Q., Yin, H., Mardani, M., Liu, M.-Y., Kautz, J., Chen, Y., and Vahdat, A. Loss-guided diffusion models for plug-and-play controllable generation. In International Conference on Machine Learning, pp.\ 32483--32498. PMLR, 2023

    2023

  46. [54]

    and Ermon, S

    Song, Y. and Ermon, S. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019

    2019

  47. [55]

    P., Kumar, A., Ermon, S., and Poole, B

    Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021 b . URL https://openreview.net/forum?id=PxTIG12RRHS

    2021

  48. [56]

    Learning generalizable perceptual representations for data-efficient no-reference image quality assessment

    Srinath, S., Mitra, S., Rao, S., and Soundararajan, R. Learning generalizable perceptual representations for data-efficient no-reference image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp.\ 22--31, 2024

    2024

  49. [57]

    Which models have perceptually-aligned gradients? an explanation via off-manifold robustness

    Srinivas, S., Bordt, S., and Lakkaraju, H. Which models have perceptually-aligned gradients? an explanation via off-manifold robustness. Advances in neural information processing systems, 36: 0 21172--21195, 2023

    2023

  50. [58]

    and Oszust, M

    Stepien, I. and Oszust, M. Tiqa-psi: Toolbox for perceptual image quality assessment of pan-sharpened images. SoftwareX, 23: 0 101494, 2023

    2023

  51. [59]

    Blindly assess image quality in the wild guided by a self-adaptive hyper network

    Su, S., Yan, Q., Zhu, Y., Zhang, C., Ge, X., Sun, J., and Zhang, Y. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3667--3676, 2020

    2020

  52. [60]

    From distortion manifold to perceptual quality: a data efficient blind image quality assessment approach

    Su, S., Yan, Q., Zhu, Y., Sun, J., and Zhang, Y. From distortion manifold to perceptual quality: a data efficient blind image quality assessment approach. Pattern Recognition, 133: 0 109047, 2023

    2023

  53. [61]

    Sddm: score-decomposed diffusion models on manifolds for unpaired image-to-image translation

    Sun, S., Wei, L., Xing, J., Jia, J., and Tian, Q. Sddm: score-decomposed diffusion models on manifolds for unpaired image-to-image translation. In International Conference on Machine Learning, pp.\ 33115--33134. PMLR, 2023

    2023

  54. [62]

    Tu, Z., Yu, X., Wang, Y., Birkbeck, N., Adsumilli, B., and Bovik, A. C. Rapique: Rapid and accurate video quality prediction of user generated content. IEEE Open Journal of Signal Processing, 2: 0 425--440, 2021

    2021

  55. [63]

    Plug-and-play diffusion features for text-driven image-to-image translation

    Tumanyan, N., Geyer, M., Bagon, S., and Dekel, T. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 1921--1930, 2023

    1921

  56. [64]

    A connection between score matching and denoising autoencoders

    Vincent, P. A connection between score matching and denoising autoencoders. Neural computation, 23 0 (7): 0 1661--1674, 2011

    2011

  57. [65]

    and Bovik, A

    Wang, Z. and Bovik, A. C. Modern image quality assessment. PhD thesis, Springer, 2006

    2006

  58. [66]

    Diffusion model-based visual compensation guidance and visual difference analysis for no-reference image quality assessment

    Wang, Z., Hu, B., Zhang, M., Li, J., Li, L., Gong, M., and Gao, X. Diffusion model-based visual compensation guidance and visual difference analysis for no-reference image quality assessment. IEEE Transactions on Image Processing, 34: 0 263--278, 2025. doi:10.1109/TIP.2024.3523800

    2025

  59. [67]

    Q-align: Teaching LMM s for visual scoring via discrete text-defined levels

    Wu, H., Zhang, Z., Zhang, W., Chen, C., Liao, L., Li, C., Gao, Y., Wang, A., Zhang, E., Sun, W., Yan, Q., Min, X., Zhai, G., and Lin, W. Q-align: Teaching LMM s for visual scoring via discrete text-defined levels. In Forty-first International Conference on Machine Learning, 20...

    2024

  60. [68]

    Z., and Shen, C

    Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., Zhou, H., Shou, M. Z., and Shen, C. Datasetdm: Synthesizing data with perception annotations using diffusion models. Advances in Neural Information Processing Systems, 36: 0 54683--54695, 2023

    2023

  61. [69]

    Open-vocabulary panoptic segmentation with text-to-image diffusion models

    Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., and De Mello, S. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 2955--2966, 2023

    2023

  62. [70]

    Boosting image quality assessment through efficient transformer adaptation with local feature enhancement

    Xu, K., Liao, L., Xiao, J., Chen, C., Wu, H., Yan, Q., and Lin, W. Boosting image quality assessment through efficient transformer adaptation with local feature enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 2662--2672, 2024

    2024

  63. [71]

    Featurenerf: Learning generalizable nerfs by distilling foundation models

    Ye, J., Wang, N., and Wang, X. Featurenerf: Learning generalizable nerfs by distilling foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 8962--8973, 2023

    2023

  64. [72]

    From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality

    Ying, Z., Niu, H., Gupta, P., Mahajan, D., Ghadiyaram, D., and Bovik, A. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3575--3585, 2020

    2020

  65. [73]

    and Korhonen, J

    You, J. and Korhonen, J. Transformer for image quality assessment. In 2021 IEEE international conference on image processing (ICIP), pp.\ 1389--1393. IEEE, 2021

    2021

  66. [74]

    Freedom: Training-free energy-guided conditional diffusion model

    Yu, J., Wang, Y., Zhao, C., Ghanem, B., and Zhang, J. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 23174--23184, 2023

    2023

  67. [75]

    Zhang, C., Zhang, C., Zhang, M., and Kweon, I. S. Text-to-image diffusion models in generative ai: A survey. arXiv preprint arXiv:2303.07909, 2023 a

    2023 arXiv

  68. [76]

    Zhang, L., Zhang, L., and Bovik, A. C. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24 0 (8): 0 2579--2591, 2015

    2015

  69. [77]

    Blind image quality assessment using a deep bilinear convolutional neural network

    Zhang, W., Ma, K., Yan, J., Deng, D., and Wang, Z. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30 0 (1): 0 36--47, 2020

    2020

  70. [78]

    Blind image quality assessment via vision-language correspondence: A multitask learning perspective

    Zhang, W., Zhai, G., Wei, Y., Yang, X., and Ma, K. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 14071--14081, 2023 b

    2023

  71. [79]

    A unified conditional framework for diffusion-based image restoration

    Zhang, Y., Shi, X., Li, D., Wang, X., Wang, J., and Li, H. A unified conditional framework for diffusion-based image restoration. Advances in Neural Information Processing Systems, 36, 2024

    2024

  72. [80]

    Unleashing text-to-image diffusion models for visual perception

    Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., and Lu, J. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 5729--5739, 2023

    2023

  73. [81]

    Metaiqa: Deep meta-learning for no-reference image quality assessment

    Zhu, H., Li, L., Wu, J., Dong, W., and Shi, G. Metaiqa: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 14143--14152, 2020

    2020

  74. [82]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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