acceptodds
Under review as a conference paper at ICLR 2027

JLD: PERCEPTUAL DISTANCE THROUGH A JACOBIAN LENS

Abstract

mage compression, restoration, and generation all require a way to measure how different two images look to a person. Pixel error ignores how people see, while the most accurate perceptual distances are typically fitted to human judgments, tying them to a fixed data and resolution. For example, when image resolution is doubled, the correlation of DISTS with human scores on TID2013 drops from 0.815 to 0.717. We introduce the Jacobian Lens Distance (JLD), which derives its perceptual geometry from a frozen vision encoder rather than from human labels. JLD combines the locality of early patch features with the perceptual sensitivity captured by later encoder representations. Specifically, we use the encoder Jaco- bian to identify directions in the early feature space that most strongly affect the encoder output, producing a fixed metric tensor, E[J⊤J], which we call the Jaco- bian lens. The lens is fitted only once from 100 unlabeled images, taking about 35 seconds. Locally, this construction defines a pullback metric in pixel space, giving JLD a clear geometric interpretation that can be directly analyzed on real images. Across four standard perceptual databases, JLD achieves state-of-the-art performance and consistently outperforms LPIPS, DISTS, PieAPP, and Dream- Sim. JLD is also robust to changes in image resolution, on TID2013, its lens-term correlation remains nearly unchanged when the resolution is doubled, decreasing only from 0.850 to 0.845. We further introduce JLD-fast, which is 4×faster than LPIPS-VGG while achieving a mean correlation of 0.911. Finally, JLD naturally extends to video, reaching a correlation of 0.786 on Waterloo IVC 4K compared with 0.611 for VMAF.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
metric, training free, Jacobian lens, gradient, frozen vision encoder, geometric distance

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_dcqAOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce the Jacobian Lens Distance (JLD), a perceptual similarity metric derived from a pretrained vision encoder. By leveraging the network's Jacobian, the authors identify which directions in the early-layer feature space most strongly influence the encoder's output and use these directions to construct a distance metric between images. The authors then demonstrate that JLD outperforms several existing perceptual metrics on established benchmarks. Moreover, unlike methods explicitly trained for perceptual similarity, JLD is resolution-agnostic while remaining computationally efficient.

Critical strengths
  1. The authors introduce a novel perceptual metric derived from a pretrained DINOv2 encoder and demonstrate that it outperforms several existing metrics on established perceptual similarity benchmarks. These results provide empirical support for the hypothesis that feature directions which strongly influence the encoder's output are also informative about human perceptual sensitivity.

  2. The authors provide an insightful geometric analysis of their metric in pixel space, illustrating how its sensitivity varies across different image distortions and regions.

  3. The paper is well written, clearly structured, and easy to understand and follow.

Critical weaknesses
  1. The authors discuss perceptual similarity broadly, without clearly distinguishing between different types of similarity. Their evaluations primarily focus on low-level distortions, such as noise, blur, and compression, whereas metrics like DreamSim target higher-level variations, including style, viewpoint, and pose. Since the paper does not evaluate these higher-level aspects of perceptual similarity, the authors should better qualify their claims of superiority over existing perceptual metrics.

  2. The authors show that existing perceptual metrics suffer from performance degradation when evaluated at resolutions different from those used during training. However, it is unclear why the images cannot simply be resized to the training resolution before computing similarity. I can see this being problematic when downsampling removes important fine details or small objects, but the authors do not investigate such cases. Without such an evaluation, the practical advantage of JLD's resolution robustness remains unclear.

Questions for authors
  1. Have you evaluated JLD on benchmarks that focus on higher-level perceptual variations, such as DreamSim's NIGHTS dataset? If not, do you expect JLD to perform similarly well on these variations, given that the current evaluations primarily focus on low-level distortions?

  2. Have you evaluated JLD on images where downsampling to a lower resolution would remove important visual information, such as fine details or small objects? Such experiments could help demonstrate the practical advantage of JLD's resolution robustness over metrics that typically operate at a fixed, lower resolution.

3 comments
AuthorsOpenReview ↗

Response to Reviewer dcqA

We sincerely thank the reviewer for the thoughtful review. We are glad that the reviewer finds the paper well written and easy to follow, the pixel-space geometric analysis insightful, and the results supportive of our hypothesis that output-influential feature directions are informative about human sensitivity. We address the questions bellow.

W1 / Q1

We thank the reviewer for this distinction. JLD is designed as a full-reference distance for low-level perceptual differences, the image-quality setting of TID2013, CSIQ, LIVE, KADID and PIPAL, and our claims are made there. Following the reviewer's suggestion we evaluated on NIGHTS: test split, 1,824 triplets, 2AFC accuracy, DreamSim's protocol (without any retraining).

MetricNIGHTS 2AFC
DreamSim (trained on NIGHTS)0.962
DINOv2-S CLS cosine0.866
DISTS0.861
JLD0.813
LPIPS-Alex0.760
LPIPS-VGG0.746

With no training on human judgements, JLD is above both LPIPS variants and below DreamSim, which is trained on NIGHTS itself. This is the ordering we expect. The lens acts on early, local patch features, so it is specialised to low-level differences. NIGHTS varies pose, layout and semantics, which is DreamSim's target. On BAPPS 2AFC, which sits between the two regimes and is LPIPS's training set, JLD scores 0.658 (note that without training) against 0.668 to 0.691 for LPIPS, DISTS and DreamSim.

Following the suggestion, we will update our mansucript with this discussion on NIGHTS dataset.

**W2 / Q2: **

We thank the reviewer for this practical question. Resizing is a reasonable remedy, and our main comparison already does it to every baseline that is release. In Table 1, DISTS, DeepDC and DeepWSD are evaluated at their released 256-pixel short side and DreamSim at 224 × 224, while JLD can run at native resolution. The Table 1 result therefore does not depend on any resolution argument.

Section 5.4 asks a different question, how each metric behaves when it is applied directly at the given resolution. On TID2013, from a 256 to a 512-pixel short side, the lens term moves from 0.850 to 0.845. LPIPS-VGG, which has no resize rule, falls from 0.757 to 0.661, and DISTS falls from 0.815 to 0.717 when its resize is not applied (with it, 0.818 as in Table 1). We will state this protocol next to the examples in our manuscript to a better read.

The advantage of JLD is that it needs no such step. Resizing to 256 pixels discards content above that resolution, which is the fine-detail case the reviewer describes, and the error map it returns is at the reduced size. JLD has no fixed input size and returns a patch-level map at native resolution, which is extremely important in modern high-resolution compression and restoration applications.

On Q2: Following the reviewer's suggestion, we ran this test on rated high-resolution content: the 120 clips of AVT-VQDB-UHD-1 test 1 (4K sources with human opinion scores), scored on 1920 × 1080 frames, 32 frames per clip. SRCC with the opinion scores:

MetricNative 1080pResized to 256
JLD0.9030.895
DISTS0.9280.854
LPIPS-VGG0.8780.837
DreamSim (released 224 × 224 input)0.875

On this content the resize costs DISTS 0.074 and LPIPS 0.041. JLD at native resolution (0.903) is above DISTS, LPIPS and DreamSim at their reduced input sizes, and each difference is significant under a paired bootstrap. JLD itself changes by only 0.008. Applied natively, DISTS is the highest here (0.928). So the preprocessing that suits DISTS depends on the content, while JLD is used one way, at native resolution, and is accurate in both cases. This is the practical advantage we claim, and we will add the experiment to the paper.

Thank you again for the constructive review. We believe the qualified claims and the NIGHTS results make the paper stronger and more useful for the ICLR community. If these answers address the concerns, we would be grateful if the reviewer would consider raising the score. We are happy to answers more questions during the discussion phase.

Reviewer_dcqAOpenReview ↗

Response to the Authors

Thanks you for addressing my comments. I will update my score accordingly. Nice job!

Dear Reviewer dcqA,

We appreciate your prompt response, and we are happy to see that you are satisfied with our rebuttal.

Reviewer_vqKVOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a novel perceptual distance called Jacobian Lens Distance (JLD), which derives its perceptual geometry from a frozen vision encoder without human label. The authors claimed that the JLD is robust to image resolutions and outperforms all learned deep distance such as LPIPS and DISTS.

Critical strengths
  1. A novel perceptual measurement method JLD is proposed with a detailed mathematical proof. The geometric properties of JLD in local shape and under different distortion scenarios are clearly presented through visualization methods.
  2. Despite sub-optimal performance compared with some Learned deep distances on the PIPAL val test set, JLD has superior performance and competitive inference time cost in the remaining test sets with synthetic distortion.
Critical weaknesses

None

Questions for authors
  1. Since JLD lags behind DISTS on PIPAL val, does it mean that its applicability to restoration algorithms and generative artifacts is limited? Specifically, can it be used as a full-reference evaluation metric for the enhancement algorithms just like LPIPS and DISTS.
  2. Although DINOv2 is trained on large-scale visual data, since the design of JLD relies on the sensitivity of the output to the early feature space, can better results be achieved by using quality-trained encoders with NR-IQA as the training target instead of it? In the main paper, lines 423-425 mentioned that frozen TOPIQ and DOVER encoder are used as reference distances in Appendix F. However, I could not locate these results. Could the authors kindly point me to the relevant section or table?
3 comments
AuthorsOpenReview ↗

Response to Reviewer vqKV

We sincerely thank the reviewer for the positive and encouraging review. We are glad that the reviewer finds our metric novel, and the geometric properties clearly visualised, and the accuracy and inference cost competitive. We answer both questions below:

Q1

Yes, JLD can be used as a full-reference metric for enhancement algorithms, in the same way as LPIPS and DISTS. To see where the gap to DISTS comes from, we split the larger PIPAL public set (23,200 pairs; none of these metrics is trained on PIPAL) by restoration class, identified from PIPAL's file naming. SRCC:

ClassPairsJLDDISTSLPIPS-VGGDeepDCPSNR
Denoising2,6000.7270.7010.7180.7230.445
SR + denoising2,8000.6270.5810.6100.6290.306
Traditional distortions5,4000.5560.5130.4030.5670.323
Traditional SR2,4000.6220.6330.5960.6580.484
PSNR-oriented SR3,2000.7170.7550.7280.7870.549
Kernel-mismatch SR2,0000.5180.5910.5610.6280.256
GAN-based SR4,8000.5000.5460.4860.5840.291
All23,2000.6030.6150.5730.6540.407

JLD is the best of the 13 metrics we scored on denoising outputs, and it is ahead of DISTS on joint SR + denoising and on traditional distortions. It is behind DISTS in some classes by only 0.01 to 0.07. Our reading is that the lens, which compares aligned early patch features, penalises such texture more than people do, while DISTS was designed to tolerate it. The output term already helps here: PIPAL val is 0.575 for the lens alone and 0.624 with it.

We will add this table and this guidance: JLD is well suited for denoising general enhancement, and it wins over LPIPS.

Q2

We thank the reviewer for reading the paper so closely, and we apologise for the pointer at lines 423 to 425. That appendix was cut while formatting the submission (iclr pressure :) ). The results are below and the appendix is updated in the revision.

Lens fitted on TOPIQ-NR (ResNet-50 trained on KonIQ-10k), SRCC:

DistanceLIVECSIQTID2013
TOPIQ-NR quality head0.8350.7480.445
TOPIQ-NR raw features0.9240.8580.673
TOPIQ-NR PCA, 64 directions0.9340.8610.685
TOPIQ-NR lens0.9270.8650.680
DINOv2-S lens (same setting, lens only)0.9420.9560.857
TOPIQ-FR (supervised reference, trained on KADID-10k scores)0.9760.9670.917

Lens fitted on the DOVER technical branch, 120 AVT-VQDB-UHD-1 clips, source-mean SRCC: lens 0.897, raw features 0.852, PCA 0.858, DOVER's own head 0.917.

The construction does transfer, it turns a no-reference model into a usable reference distance. However, a quality-trained encoder did not give a better lens than DINOv2. The lens gains less than 0.01 over that encoder's raw features, and the self-supervised DINOv2 lens is far ahead (0.857). Our 14-encoder study points the same way, every self-supervised encoder gives a better lens than every supervised or language-supervised one (DINOv2-S 0.913 on development data, CLIP-B 0.828, VGG16 0.813, ConvNeXt-B 0.765). Our hypothesis is that a scalar quality output yields a far less informative Jacobian than a high-dimensional self-supervised embedding. This supports the choice of a self-supervised encoder in the paper.

TOPIQ-FR is listed as a supervised reference. It is regressed on KADID-10k opinion scores and is higher on TID2013 and LIVE, as expected for a model trained on such labels. JLD reaches 0.877, 0.971 and 0.964 on the same sets with no opinion scores in its fit, and it has the highest four-set mean among the 15 distances of Table 1, which include the human-fitted LPIPS, DISTS and PieAPP.

Thank you again for the encouraging review. We were glad to see the geometric analysis and the label-free result valued, and we hope JLD proves useful to the ICLR community as a perceptual distance that needs no human labels to fit. We are happy to answer any more questions during the discussion phase.

Thanks for the additional experiments. Although the SR methods used in PIPAL are somewhat outdated—leaving room for future FR metrics to better accommodate modern generative ISR methods—this does not diminish the novelty of the work. Thus, I keep my rating unchanged.

Dear Reviewer vqKV,

We thank you for your response and appreciation of our work.

Reviewer_5RCmOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose the study of the sensitivity of a model to image distortions by analyzing its Jacobian (in-text Eq. in the abstract and Eq. 1), which is equivalent to measuring distances in the feature space using the Euclidean distance. This of course, implies a non-euclidean metric in the spatial domain directly related to the Jacobian: M=J^T*J. This metric can include more or less behavior of the model by controlling the number of coefficients of an eigendecomposition of this metric (d in Eq.1). The authors make a thorough qualitative analysis of the non-Euclidean nature of this metric (section 4 and fig.2). Moreover, they propose a final distance, D, in the spatial domain which combines the above concept M (and its eigendecomposition) with a CSF prefilter and a sort of cosine similarity in the feature space (Eq. 2) and linearly approximate this combination to put in the standard quadratic-metric form (Eq. 4). The authors compare this final distance with a number of other metrics and they find it overperform those in terms of correlation with human opinion on a number of subjectively-rated databases of distorted images (section 5.1-5.3). They also find that this distance is more insensitive than conventional metrics to changes in resolution (section 5.4).

Critical strengths
  • S1 - The main concept is geometrically appealing and correct. While it has been already applied to artificial vision models (see W1 below for examples), it is important to show more instances of this concept in the ICLR community.

  • S2 - The experiments studying alignment with human opinion are extensive and properly point out the usefulness of the proposed distance in the image quality problem.

  • S3 - Extensive appendices include geometrical proves, and interesting perceptual experiments like obtaining the Contrast Sensitivity Function of the system and analysis of group-MAximum Differntiation.

Critical weaknesses
  • W1 - The main geometrical concept is interesting but not new. While the authors cite classical works where metrics under change-of-basis lead to the basic result in section 2, M=J^T*J (assuming Euclidean metric in the inner domain of the model), they miss some references there (easy to solve problem ;-). More importantly, the authors have missed that (i) this geometrical concept has been applied to other artificial vision models before, and (ii) the use of Gaussian noise as probe to obtain the geometry (the shape of the ellipsoids in the input and in the inner domain) has a long tradition too. I give some examples of these references below.

  • W2 - The mixture of concepts (M=J^T*J, CSF prefilter, and cosine similarity) to propose the final distance D is not well justified in the text. (To me) Eq. 3 seems to come from no-where, or better said, from a principle which is in contradiction with the main spirit of the work. Let me explain: while the main story is focused on the (geometrically well-founded) jacobian of the model and the use of no human vision information, the proposed distance "suddenly" uses human-like CSF pre-filters and combines the jacobian metric with cosine similarity. I understand that authors wanted to raise the correlation with humans from 0.905 to 0.926 in Fig.6 for the particular image quality application, but (to me) this breaks the nicest story, isnt it?.

  • W3 - In the abstract and in section 5.4 authors are concerned about the robustness of the distance measurements under changes of resolution. In my view this cannot be addressed using the MOS in the considered databases because it should certainly change if the images are resampled and re-rated by humans, and this information is not available in those databases.

Questions for authors

***Related to W1:

** Please consider citing/diacussing these papers (that use M in human vision models):

  • Pons et al. (1999) Image quality metric based on multidimensional contrast perception models Displays, 20(2):93-110, 1999 (first use of the change of coordinates and euclidean assumption to get M = J^T)

  • da Fonseca, M., & Samengo, I. (2016). Derivation of human chromatic discrimination ability from an information-theoretical notion of distance in color space. Neural Computation, 28(12), 2628–2655 (first use of an information theoretic Jacobian to propose the metric M before Berardino et al. 2017)

  • da Fonseca & M Samengo I (2019) Novel perceptually uniform chromatic space Neural Computation 10.1162/neco_a_0107330:6 (1612-1623) (Use of this metric and connection with the change of coordinates cited above and local statistics of the signal)

** Please consider citing/discussing these papers (that look at the Jacobian of deep-nets to assess their sensitivity)

  • A Gomez-Villa et al. (2020) Color illusions also deceive CNNs for low-level vision tasks: Analysis and implications. Vision Research 176, 156-174 (first to analyze the Jacobian to explain CSF-like sensitivity of deep-nets)

  • Li et al. (2022) Contrast sensitivity functions in autoencoders J. Vision 2022, Vol.22, 8. doi:https://doi.org/10.1167/jov.22.6.8 (different analysis of the linearization -Jacobian and average linear approximations over many images- to point out the CSF-like sensitivity of deep-nets)

** Please consider citing/discussing these papers (that analyze the transform of noise along vision models and its relation with the Jacobian -as in your discussion in Section 3.2 and Proposition 2-)

  • Ahumada, A. J. (1987). Putting the visual system noise back in the picture. JOSA A, 4(12), 2372–2378. (first to put the inner visual noise in the spatial domain, analyzing the transforms of noise in terms of the jacobian)

  • Malo et al. (2025) Estimating the contribution of early and late noise in vision from psychophysical data. J. Vision 2025, Vol.25, 12. doi:https://doi.org/10.1167/jov.25.1.12 (joint use of noise propagation and distance metrics together with the jacobian of the vision model)

***Related to S3 and W2 (experiment with gratings):

I fully support the concern of the authors about the behavior of the model with gratings because it is the classical characterization of human behavior in vision science (now studied in CNNs -see references above-, transformers and LLMs -Hernández-Cámara et al. (2026) Contrast sensitivity in multimodal large language models Neural Nets, Volume 201, 2026-). However, what the reviewers find in their interesting Fig. A14 is similar to some results of Li et al. JoV 2022 and Akbarinia et al. Neural Nets 2023 and Hernandez-Camara et al. 26: the plain model DOES NOT exactly reproduce the CSF, and only does it better if the (by-hand included) pre-filter is considered. This is related to my criticism above about breaking the main logic of the work. However, I acknowledge I was interested in seing this experiment (despite the fact that the ablation result show that including the CSFs does not a big improvement)... which speaks well about the properties of the selected model (and intrinsic its Jacobian).

*** Related to W3

If I didnt miss any comment in line with what I said in W3 (which I may have, excuses! -ICLR reviewing pressure-), I would remove mention to this in the abstract and the whole section 5.3. I agree that dependence of distance with resolution is an important avenue of research, but one cannot assume that the MOS values of TID and KADID should be invariant. Other literature may provide human data on this, but I guess MOS depends on resolution. Nevertheless, I think that removing mentions to resolution does not reduce the quality of the study.

*** OVERALL COMMENT: I am happy to consider a rise in the score if these questions are addressed because tha anlysis of the sensitivity and the Jacobian of artificial vison models is important for the ICLR community and has parallelism with previous studies in human vision

4 comments
AuthorsOpenReview ↗

Response to Reviewer 5RCm

We sincerely thank the reviewer for the careful review. We are glad that the reviewer finds the main concept geometrically appealing and correct, the human-opinion experiments extensive and the appendices valuable. We also appreciate the references from the human-vision literature. We address each point below.

W1

We thank the reviewer for highlighting these references. We will cite and discuss them in our updated manuscript.

W2

We thank the reviewer for this point. The Jacobian remains the core of the method, and we will improve the presentation so that this is clear. The lens is the main source of the result. i.e:

VariantTID2013CSIQLIVEKADIDMean
Raw block-1 features0.6600.8410.8590.7270.772
PCA, 64 directions0.7660.9120.9250.8000.851
Lens only (Eq. 1)0.8570.9560.9420.8660.905
Lens + output term, no pre-filter0.8560.9700.9610.8880.919
Lens + pre-filter (JLD-fast)0.8750.9560.9450.8700.911
Full JLD (Eq. 3)0.8770.9710.9640.8920.926

The lens accounts for 0.133 of the 0.154 gain over raw features. With nothing added it would place third against the 15 baselines of Table 1, behind VSI (0.915) and DeepWSD (0.913) and ahead of DISTS (0.900), PieAPP (0.881) and DreamSim (0.871).

The output term comes from the same encoder. The lens is the first-order sensitivity of the encoder output to early features, resolved per patch. The output term, measures that same output directly for a finite displacement, and it covers whole-image changes such as a global colour shift, where the patch-averaged lens is weak. It is second order near the reference, so the pullback geometry of Eq. 4 is unchanged. And we agree with the reviewer that in a absolute sense it breaks a nice story around geometry in flow of our paper. But we share our intuition plainly here. As the reviewer observes, it raises the correlation; that choice was made on the development sets, taking the model properties into considerstaion to include the pre-filter. which we shall state in our limitations. Lens + output term, with no human-vision component at all, reaches 0.919, level with the best baseline (VSI 0.915).

Eq. 2 low-passes the two chroma channels only. It is the one component not derived from the encoder, and it adds less than 0.01. As the reviewer notes, this small gain speaks well of the selected model and its intrinsic Jacobian. Our grating experiment agrees with the reviewer's reading of Fig. A14 and with Li et al. and Akbarinia et al. The lens already moves the encoder toward the human chromatic CSF: shape correlation with the human curve is 0.72 (red-green) and 0.56 (blue-yellow), against 0.09 and −0.23 for raw features, and the pre-filter closes part of the remaining gap (0.84 and 0.72). For luminance the lens does not reproduce the human CSF (0.42, against 0.98 for raw features). We will discuss this in the paper for better clairty.

W3

We thank the reviewer for this important point. We agree that the scores were collected under one viewing condition and the resampled images were not re-rated. We will rewrite our Section 5.4 to discuss this point more clearly and will not over emphasise it as a claim. When every image of a database is resampled by the same factor, the relative ordering of its distorted images should stay almost the same, so a metric that does not depend on input size should keep its agreement with the original scores. The data support this assumption: on TID2013, from a 256 to a 512-pixel short side, PSNR moves from 0.712 to 0.713, NLPD stays at 0.787 and VSI moves from 0.887 to 0.888. Under the same test the lens term moves from 0.850 to 0.845, while LPIPS-VGG falls from 0.757 to 0.661 and LPIPS-Alex from 0.824 to 0.697. JLD has no fixed input size, whereas the weights of earlier deep metrics were fitted at one resolution. We will reword the abstract and Section 5.4 to state this assumption and to describe the result as stability of the metric under uniform resampling.

Thank you again for the encouraging and constructive review. We are glad that the reviewer considers the analysis of the sensitivity and the Jacobian of artificial vision models important for the ICLR community. If these answers address the concerns, we would be grateful if the reviewer would consider raising the score. We are happy to engage in more discussions.

AuthorsOpenReview ↗

Gentle Reminder: Discussion period ending soon

Dear Reviewer,

Just a brief reminder that the Discussion period concludes in a few days.

Thank you again for your thoughtful and thorough reviews. We provided a detailed response to all your concerns in our rebuttal. If you feel we addressed your concerns, we would really appreciate it if you could acknowledge and consider reassessing your score. We would be happy to address any remaining concerns during the remainder of the discussion period.

Thank you again for the time and effort you put into the review.

AuthorsOpenReview ↗

Gentle Reminder: Discussion period ending soon

Dear Reviewer 5RCm,

Just a brief reminder that the Discussion period concludes soon, we would very much like to know if you were satisfied with our rebuttal. We are looking forward to your response.

Thank you again for the time and effort you put into the review.

AuthorsOpenReview ↗

Response Appreciated - Discussion Phase Ends Today

Dear Reviewer 5RCm,

Reminder that the Discussion period concludes today, and we are eager to hear your response to our rebuttal. We would very much like to know if you were satisfied with our rebuttal. We are looking forward to your response.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Comment

We sincerely thank the three reviewers for their careful and constructive reviews. We are encouraged that the reviewers find the main concept geometrically appealing and correct (5RCm), the metric novel with detailed proofs and clear visualisation (vqKV), and the paper well written with an insightful geometric analysis (dcqA), and that Reviewer 5RCm considers this line of analysis extremely important for the ICLR community.

We answer every weakness and question in the individual responses. In summary, we provide:

  1. Lens-only ablation (5RCm). The Jacobian lens alone reaches 0.905 four-set SRCC, and lens + output term reaches 0.919 with no human-vision component.
  2. Prior work (5RCm). We add and discuss the vision-science and deep-network Jacobian literature the reviewer lists.
  3. Resolution (5RCm, dcqA). We clarify what the resolution study measures and the protocol behind it.
  4. Scope and NIGHTS (dcqA). We add NIGHTS results and clarify the distinction of application.
  5. PIPAL by restoration class, and quality-trained encoders (vqKV). We add both sets of results, including the TOPIQ and DOVER tables.

We would be glad to provide further answers and clarifications during the discussion phase.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.