Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

ShapeMark: Robust and Diversity-Preserving Watermarking for Diffusion Models

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read ShapeMark encodes watermark bits in noise permutations, not values, and claims near-perfect recovery under heavy distortion while preserving generation diversity.

desk verdict ShapeMark is a genuinely new structural-encoding NaW method with a clean diversity fix, but the extreme drop/crop robustness numbers need independent validation before I'd trust them. read the letter →

arxiv 2603.09454 v1 pith:7NVCDYAK submitted 2026-03-10 cs.CR

Yuqi Qian , Yun Cao , Haocheng Fu , Meiyang Lv , Meineng Zhu This is my paper
classification cs.CR
keywords diffusionmodelwatermarkingnoise-as-watermarkstructuralencodingpermutationcodebookDDIMinversiongenerationdiversitycontentprovenancerobustness
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

Diffusion-generated images lack a built-in proof of origin, and the usual fix—hiding bits in the initial noise and recovering them by inverting the generation—fails in practice because recompression, cropping, and blur corrupt individual noise values. ShapeMark argues that this fragility is a design choice, not a law: prior noise-as-watermark methods tie each bit to a value, so a small perturbation near a decision boundary flips the bit. ShapeMark instead embeds bits in the relative arrangement of noise elements—the latent is sorted by magnitude into separable bins, carved into blocks, and a payload-chosen permutation moves blocks between group slots—so decoding asks which block moved where, a question that survives value-level noise. A second, exactly invertible shuffling step randomizes where blocks sit in each image, so the payload does not fix a spatial pattern and generation diversity holds. If the claim is right, provenance and traceability for AI imagery become practical even after the heavy edits that images typically undergo.

What carries the argument

Structural Encoding (SE): a keyed template over the noise latent plus a payload-conditioned rearrangement. Latent elements are ranked by magnitude and split into Q quantile bins so blocks from different bins are statistically separable; a key-seeded shuffle decides which coordinates bundle into each block, and groups align one block per bin. Each 4-bit payload chunk selects a permutation from a position-balanced 16-codeword codebook (16 of the 24 permutations of 4 blocks) and moves block values between group slots. Decoding regenerates the canonical latent from the key, reads observed blocks, and picks the codeword minimizing summed L2 distance to reference blocks; a normalized margin betwee

What would settle it

Two checks. (1) For each of the nine distortions, measure the empirical 'block-flip rate'—the fraction of recovered latent blocks whose quantile-bin membership differs from the key-defined template—and plot it against per-bit accuracy; the claim predicts they track each other, while accuracy collapsing with few flips would implicate a different mechanism. (2) Run the full verification pipeline on held-out non-watermarked images and on images watermarked under different keys, counting how often the detection statistic exceeds the calibrated threshold; if the empirical FPR is an order of magnitu

Watch

Extended reading notes

Core claim

Central claim: watermark robustness in diffusion models depends not on how strongly bits are stamped into noise values but on how decoding is posed. ShapeMark poses it as permutation recovery: sort the initial latent by magnitude into separable quantile bins, bundle elements into key-ordered blocks, group one block per bin, and let each 4-bit payload chunk pick a permutation from a balanced 16-codeword codebook moving blocks between slots. Decoding inverts the image, regenerates the canonical latent from the key, and matches blocks by L2 distance; the winning margin feeds detection. Perturbations alter individual values but rarely reorder whole blocks across magnitude regions, so the permuta

Load-bearing premise

The paper's robustness argument assumes that DDIM inversion with an empty prompt and only ten steps returns a latent whose block-level magnitude ordering is close enough to the original that codebook matching still succeeds after lossy attacks; if inversion error shuffles blocks across quantile boundaries, decoding fails.

Editorial extensions

If this is right

  • A 256-bit payload (user identity, session metadata) can be bound to each generated image and recovered from a weakly conditioned inversion that uses an empty prompt and only 10 steps, so verification does not require the original prompt.
  • The claimed robustness covers nine practical distortions—JPEG, resize, random drop, random crop, two blurs, two noise types, and brightness—with per-bit accuracy above 0.93 in every reported case.
  • Robustness and diversity are separated into two mechanisms (SE and PDSR), so engineers can tune capacity, robustness, and diversity along relatively independent axes instead of facing a single trade-off.
  • Because embedding is a pure value-preserving permutation of a key-sampled Gaussian and no model weights change, the scheme transfers across diffusion backbones and deployments without retraining.
  • At 2048-bit capacity the method still recovers about 85% of bits under attack, which extends the approach beyond binary provenance to fine-grained, high-cardinality attribution.

Reading between the lines

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

  • If robustness is really carried by block-level ordering, bit accuracy should degrade exactly when inversion error moves blocks across quantile boundaries. A per-distortion plot of block-flip rate versus bit accuracy—absent from the paper—would isolate the failure mechanism and predict attack ceilings.
  • The near-perfect bit accuracy under 80% random drop and 60% random crop (0.9985 and 0.9977) is stronger than the paper's 'ordering survives' intuition suggests; either those attacks leave the latent's dominant structure intact after inversion, or the group-margin statistic is harvesting more redundancy than the paper accounts for.
  • The FPR = 10^-6 threshold used throughout is extrapolated from 10,000 null images via a generalized Pareto tail fit, a step the appendix flags as needing larger held-out and cross-key validation. If the true null tail is heavier, reported TPRs at that FPR would shift.
  • The codebook deliberately uses 16 of the 24 permutations of four blocks—a rate reduction the paper likens to channel coding. That framing invites an obvious extension: error-correcting codes across groups, or larger codebooks, to push robustness toward geometric attacks that destroy block alignment.
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

3 major / 5 minor

Summary. ShapeMark proposes a noise-as-watermark (NaW) method for diffusion models that encodes a payload in the relative ordering of key-selected quantile blocks of the initial noise latent, rather than in individual noise values. Structural Encoding (SE) builds blocks and groups from a key-generated template of the canonical Gaussian latent and embeds bits via group-wise block permutations; Payload-Debiasing Structural Randomization (PDSR) reshuffles block positions in a payload-agnostic, exactly invertible way to preserve generation diversity. Decoding uses DDIM inversion of the queried image, undoes PDSR with a public nonce, and performs codebook matching against the key-regenerated canonical latent; a margin-based detection statistic is calibrated to a target FPR. Experiments on Stable Diffusion v2.1 report state-of-the-art robustness (TPR 0.999 at FPR=10^-6 under attack, per-bit accuracy 0.9870) while achieving the highest LPIPS diversity (0.7338). The paper includes detailed algorithms, an ablation study, and per-distortion bit-accuracy tables.

Significance. If the reported numbers are reproducible, ShapeMark would be a meaningful advance: it provides a distribution-preserving, keyed, payload-capable watermarking scheme whose robustness derives from group-level structural relationships rather than fragile value constraints, and it explicitly targets the diversity loss typical of repeated NaW embedding. The design is coherent, and the ablation study (Table 2) supports the complementary roles of SE (robustness) and PDSR (diversity/quality). The explicit algorithms and the balanced, canonical codebook (Appendix A) are useful contributions. However, the central empirical claims—especially the extreme robustness under Drop/Crop and the FPR=10^-6 operating point—are not yet supported by enough technical evidence, so the paper needs additional validation before its headline claims can be accepted.

major comments (3)
  1. [Table 5 / Sec. 3.5] The robustness numbers under Random Drop (80% area) and Random Crop (60% area) are implausibly high without an analysis of the recovered latent. Bit accuracies of 0.9985 and 0.9977 under attacks that destroy 80% and 40% of pixels, respectively, imply that DDIM inversion plus block-ordering decoding recovers almost perfectly from images that are largely zeroed. The paper provides aggregate TPR/bit accuracy but no per-distortion measurement of how close the inverted latent is to the true latent, nor the fraction of groups for which the correct permutation is the argmin. I recommend adding a latent-recovery analysis (e.g., L2 distance between inverted and true latent, permutation recovery rate per distortion, and a comparison of the margin statistic under Drop/Crop against clean) and clarifying the exact attack implementation (e.g., whether the dropped/cropped regions are zero-filled, how t
  2. [Appendix B.1, Eqs. (24)-(26)] The detection threshold for FPR=10^-6 is obtained by Generalized Pareto Distribution tail extrapolation from n=10,000 null scores, which provide empirical support only down to 10^-4. The reported TPR at FPR=10^-6 therefore rests entirely on the validity of the GPD tail model. The paper itself notes in Appendix B.1 that deployment-grade calibration would require larger null sets and cross-key validation. If the tail model is even slightly optimistic, the headline TPR=0.999 at FPR=10^-6 is inflated. I ask for a sensitivity analysis (e.g., compare GPD-extrapolated thresholds against larger null sets, cross-key nulls, or nonparametric bounds; report TPR at the empirically verifiable FPR=10^-4 as well). This is load-bearing for the detection claim.
  3. [Table 1 / Sec. 4.1] The headline TPR and per-bit accuracy numbers are point estimates without confidence intervals. Given that the paper reports standard errors for CLIP and FID and uses t-tests for those metrics, it is inconsistent to omit uncertainty quantification for the central robustness and traceability results. I recommend reporting bootstrap confidence intervals for TPR and bit accuracy (and, where applicable, for the difference against the strongest baseline, T2SMark). Without this, it is hard to judge whether ShapeMark's 0.999 vs. 0.998 TPR and 0.9870 vs. 0.9859 bit accuracy are materially significant.
minor comments (5)
  1. [Sec. 3.3, Eq. (7)] The notation z(e)[I_{j,g}] = z[I_{sigma_g(j),g}] is slightly confusing because the destination block and source block are both indexed by the same group; a short example or a comment clarifying that I_{j,g} are destination slots would help.
  2. [Fig. 2 caption] The caption says 'each block is shown with a block size of one' but the actual blocks in Figs. 2 and 3 appear as colored regions. This is fine, but the caption should be clearer that the figure is schematic and not to scale.
  3. [Table 5] The row label 'Adv.(ave)' should be 'Att.(ave)' (or 'Attack average') for consistency with the text.
  4. [Sec. 4.3, Table 3] The results for inversion steps 50 and 100 are slightly worse than for 10 steps at most guidance scales. The text attributes increasing steps to 'more accurate recovery,' which is not supported by the table; either correct the interpretation or explain the non-monotonicity.
  5. [Sec. 3.4] The public nonce r is said to be 'obtained from a provenance record or generation log.' In practice, if the nonce is not embedded in the image, the verification party must have access to that side information. The paper should state explicitly whether the nonce is transmitted as metadata, which may be stripped by an attacker, or whether the method is otherwise robust to nonce loss.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ShapeMark's decoding is a private-key codebook matching scheme whose reference latent is independent of the fitted threshold; robustness is an empirical claim, not a construction.

full rationale

ShapeMark's pipeline is a private-key codebook-matching watermark: the embedder permutes key-derived noise blocks (Eq. 7) and the decoder regenerates the same key-derived reference latent, undoes PDSR (Eq. 10), and selects the codebook permutation minimizing Eq. (15). This is a standard matched-filter/nearest-codeword decoder; the reference is not fitted to the queried image, and no equation reduces to a fitted quantity. The detection threshold tau_det is calibrated on 10,000 null images via GPD tail extrapolation (Appendix B.1) and then applied to watermarked images; although the paper itself notes the threshold depends on the tail fit, this is a disclosed statistical calibration, not a circular prediction. No load-bearing self-citation appears: the only supporting citations for inversion closeness point to external prior work, and the comparative baselines are external. The robustness and diversity results are empirical claims evaluated against external datasets and baselines; the plausibility of high bit-accuracy under destructive attacks is a validation/attack-implementation concern, not a circularity. Therefore no specific circular step can be quoted.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a handful of hyperparameters, a statistical tail-extrapolation threshold, and assumptions about DDIM inversion fidelity and quantile separability. No new physical or algorithmic entities are introduced beyond the keyed permutation template and the public nonce.

free parameters (6)
  • Q (number of quantile bins) = 4
    Hand-chosen default; controls group size and robustness/payload trade-off.
  • b (block size) = 64 latent elements
    Hand-chosen; larger b increases robustness per bit, smaller b increases payload capacity.
  • s (group size) = 4 = Q
    Default group size; one block per bin.
  • k (payload bits per group) = 4
    log2|C| = 4, codebook size 16.
  • balanced codebook C = 16 permutations (Table 4)
    Design choice; not all 24 S4 permutations are used to keep fixed-length coding and position balance.
  • detection threshold tau_det = 0.005542
    Calibrated from 10,000 null images via GPD tail extrapolation to target FPR=10^-6; depends on fitted GPD shape/scale parameters.
assumptions (5)
  • domain assumption DDIM inversion approximates the original initial noise latent sufficiently closely for block-level matching.
    Invoked in Sec 2.1 and Sec 3.5 Stage I; without this, decoding cannot work.
  • domain assumption Magnitude-quantile block identities remain separable after inversion error and image distortions.
    Sec 3.3 'Why SE is robust' relies on this; payload is decoded by matching block norms.
  • domain assumption A key-seeded fixed canonical latent plus random block permutation preserves the diversity of fresh Gaussian noise.
    Sec 3.4 PDSR; if a fixed multiset of values measurably reduces diversity in deployment, the diversity claim fails.
  • domain assumption GPD tail extrapolation from 10k null samples accurately estimates the 1e-6 quantile of the null score distribution.
    Appendix B.1; used to report TPR at FPR=10^-6.
  • domain assumption Attacker has access to only one final image and cannot perform collusion or multi-image analysis.
    Threat model in Sec 3.1; the fixed canonical latent is vulnerable to multi-image matching attacks not considered.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ShapeMark: Robust and Diversity-Preserving Watermarking for Diffusion Models." pith.science (2026). https://pith.science/paper/7NVCDYAK

@misc{pith2026260309454,
  author       = {Pith},
  title        = {Pith review of: ShapeMark: Robust and Diversity-Preserving Watermarking for Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7NVCDYAK}},
  note         = {Machine review of arXiv:2603.09454}
}
read the original abstract

Diffusion models have made substantial advances in recent years, enabling high-quality image synthesis; however, the widespread dissemination and reuse of their outputs have introduced new challenges in intellectual property protection and content provenance. Image watermarking offers a solution to these challenges, and recent work has increasingly explored Noise-as-Watermark (NaW) approaches that integrate watermarking directly into the diffusion process. However, existing NaW methods fail to balance robustness and diversity. We attribute this weakness to value encoding, which encodes watermark bits into individual sampled values. It is extremely fragile in practical application scenarios. To address this, we encode watermark bits into the structured noise pattern, so that the watermark is preserved even when individual values are perturbed. To further ensure generation diversity, we introduce a dedicated randomization design that reshuffles the positions of noise elements without changing their values, preventing the watermark from inducing fixed noise patterns or spatial locations. Extensive experiments demonstrate that our method achieves state-of-the-art robustness while maintaining high generation quality across a wide range of lossy scenarios.

Figures

Figures reproduced from arXiv: 2603.09454 by the authors.

Figure 1
Figure 1. Comparison between existing Noise-as-Watermark meth￾ods and the proposed method. As diffusion-based generative models are increasingly adopted in practical AIGC pipelines, the need to protect the copyright of generated images and to enable the detection of unauthorized reuse has become increasingly pressing, espe￾cially as generated content is widely disseminated and rou￾tinely subjected to lossy post-processing in … view at source ↗
Figure 2
Figure 2. Overall pipeline of ShapeMark. The figure illustrates structural-level watermark encoding in the diffusion noise latent and structure-based watermark decoding via diffusion inversion and reference codebook matching. For clarity of visualization, each block is shown with a block size of one (i.e., a single noise element per block), whereas the actual method supports larger block sizes in practice. each constituent bl… view at source ↗
Figure 3
Figure 3. The watermarked image is attacked by different distortions. (a) Watermarked image. (b) JPEG, QF = 25. (c) 80% area Random Drop. (d) 60% area Random Crop. (e) 25% Resize and restore. (f) Gaussian Blur, r = 4. (g) Median Blur, k = 7. (h) Gaussian Noise, µ = 0, σ = 0.05. (i) Salt and Pepper Noise, p = 0.05. (j) Brightness, factor = 6 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Ablation Studies. noise-as-watermark methods have reported pronounced per￾formance degradation under Gaussian perturbations. We attribute this robustness to ShapeMark’s structure encoding: while Gaussian noise perturbs individual latent values, the within-group block o…
Figure 5
Figure 5. Figure 5: Qualitative examples of watermarked images generated from diverse text prompts. The images are generated using the following prompts: (1) “A close-up photo of a golden retriever dog sitting on grass, natural light, shallow depth of field”; (2) “A bowl of fresh fruits o…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. AnchorMark: Robust Diffusion Watermarking via Latent-Space Rotation Synchrony

    cs.CR 2026-07 conditional novelty 6.0 of 10

    Image rotation is shown to rotate the recovered latent by the same angle, and a central phase anchor exploits this to estimate and undo rotation before decoding the watermark.

Reference graph

Works this paper leans on

24 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Ci, H., Yang, P., Song, Y., and Shou, M. Z. Ringid: Rethinking tree-ring watermarking for enhanced multi-key identification. In European Conference on Computer Vision, pp.\ 338--354. Springer, 2024

    2024

  2. [2]

    The stable signature: Rooting watermarks in latent diffusion models

    Fernandez, P., Couairon, G., J \'e gou, H., Douze, M., and Furon, T. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 22466--22477, 2023

    2023

  3. [3]

    An undetectable watermark for generative image models

    Gunn, S., Zhao, X., and Song, D. An undetectable watermark for generative image models. arXiv preprint arXiv:2410.07369, 2024

    arXiv 2024

  4. [4]

    Stable diffusion prompts dataset

    Gustavosta . Stable diffusion prompts dataset. https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts, 2022. Accessed: 2026-01-23

    2022

  5. [5]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017

    2017

  6. [6]

    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

  7. [7]

    Robin: Robust and invisible watermarks for diffusion models with adversarial optimization

    Huang, H., Wu, Y., and Wang, Q. Robin: Robust and invisible watermarks for diffusion models with adversarial optimization. Advances in Neural Information Processing Systems, 37: 0 3937--3963, 2024

    2024

  8. [8]

    Digital watermarking and steganography, 2008

    Ingemar, C., Matthew, M., Jeffrey, B., Jessica, F., and Ton, K. Digital watermarking and steganography, 2008

    2008

Show all 24 references
  1. [9]

    Mbrs: Enhancing robustness of dnn-based watermarking by mini-batch of real and simulated jpeg compression

    Jia, Z., Fang, H., and Zhang, W. Mbrs: Enhancing robustness of dnn-based watermarking by mini-batch of real and simulated jpeg compression. In Proceedings of the 29th ACM international conference on multimedia, pp.\ 41--49, 2021

    2021

  2. [10]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  3. [11]

    Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll \'a r, P., and Zitnick, C. L. Microsoft coco: Common objects in context. In European conference on computer vision, pp.\ 740--755. Springer, 2014

    2014

  4. [12]

    Towards blind watermarking: Combining invertible and non-invertible mechanisms

    Ma, R., Guo, M., Hou, Y., Yang, F., Li, Y., Jia, H., and Xie, X. Towards blind watermarking: Combining invertible and non-invertible mechanisms. In Proceedings of the 30th ACM International Conference on Multimedia, pp.\ 1532--1542, 2022

    2022

  5. [13]

    A watermark-conditioned diffusion model for ip protection

    Min, R., Li, S., Chen, H., and Cheng, M. A watermark-conditioned diffusion model for ip protection. In European Conference on Computer Vision, pp.\ 104--120. Springer, 2024

    2024

  6. [14]

    and Xie, S

    Peebles, W. and Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 4195--4205, 2023

    2023

  7. [15]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    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 International conference on machine learning, pp.\ 8748--8763. PmLR, 2021

    2021

  8. [16]

    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

  9. [17]

    Denoising diffusion implicit models

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  10. [18]

    Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust

    Wen, Y., Kirchenbauer, J., Geiping, J., and Goldstein, T. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023

    2023 arXiv

  11. [19]

    T2smark: Balancing robustness and diversity in noise-as-watermark for diffusion models

    Yang, J., Fang, H., Zhang, W., Yu, N., and Chen, K. T2smark: Balancing robustness and diversity in noise-as-watermark for diffusion models. arXiv preprint arXiv:2510.22366, 2025

    2025

  12. [20]

    Gaussian shading: Provable performance-lossless image watermarking for diffusion models

    Yang, Z., Zeng, K., Chen, K., Fang, H., Zhang, W., and Yu, N. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12162--12171, 2024

    2024

  13. [21]

    A., Xu, L., Cuesta-Infante, A., and Veeramachaneni, K

    Zhang, K. A., Xu, L., Cuesta-Infante, A., and Veeramachaneni, K. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019

    1909 arXiv

  14. [22]

    V., Bearfield, C

    Zhang, L., Liu, X., Martin, A. V., Bearfield, C. X., Brun, Y., and Guan, H. Attack-resilient image watermarking using stable diffusion. Advances in Neural Information Processing Systems, 37: 0 38480--38507, 2024

    2024

  15. [23]

    A., Shechtman, E., and Wang, O

    Zhang, R., Isola, P., Efros, A. A., Shechtman, E., and Wang, O. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 586--595, 2018

    2018

  16. [24]

    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 2, 2026 · model on record in the stance chip above.