Pith. sign in

REVIEW 4 major objections 6 minor 26 references

Task-Adaptive Semantic Communications with Controllable Diffusion-based Data Regeneration

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A three-phase feedback loop lets a diffusion model regenerate images that match a receiver's current task from a coarse segmentation map plus sparse attention-masked edge details.

desk verdict Genuinely new three-phase task-adaptive semantic communication architecture; the core empirical claim holds up better than the stress-test suggests, but the paper needs a matched-bit control and more reporting rigor. read the letter →

arxiv 2505.07980 v1 pith:SG6C5VGY submitted 2025-05-12 cs.CL

Fupei Guo , Achintha Wijesinghe , Songyang Zhang , Zhi Ding This is my paper · ORCID
classification cs.CL
keywords semanticcommunicationstask-adaptivetransmissiondiffusionmodelsattentionmapsCLIPclassactivationedgeimageregeneration
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

This paper proposes a three-phase semantic communication scheme in which the transmitter first sends only a coarse segmentation map, the receiver uses a diffusion model to regenerate a rough image, the receiver then tells the transmitter which objects or regions matter (as a class label or a free-text prompt), and the transmitter replies with just the edge details inside that region of interest. The claim is that this adaptive loop delivers the information a downstream task actually needs, such as object counting or detection for cars or people, or depth estimation, while transmitting far less than the full image or even a full edge map. If correct, it means one compressed semantic message can serve receivers whose tasks change over time, without retransmitting the whole image. The paper supports this with Cityscapes experiments comparing task metrics and compression rates against GESCO and Diff-GO.

What carries the argument

The mechanism that carries the argument is the task-adaptive semantic update. An attention map $A$ is produced from receiver feedback, using a class activation map for known labels or a CLIP text-image similarity map for free-form prompts, and thresholded into a binary mask $M_A=\mathbb{I}(A>\tau)$. The mask selects regions of the edge map $x_{edge}$ to form $x_{att}=M_A\odot x_{edge}$, and only the non-zero patches of $x_{att}$ plus their location information are transmitted. These sparse patches become an additional condition in a denoising diffusion probabilistic model $p_\theta(x_0|x_{seg},x_{att})$, so the receiver's request for more detail is converted into a small, task-localized semantic update instead of a new image transmission.

What would settle it

Inject controlled bit errors or Gaussian noise into the transmitted segmentation map, edge-patch latent, or feedback text in a simulated channel, rerun the Cityscapes object-detection and depth evaluations, and compare object-count MSE and mIoU against the lossless results in Table II. If a small error rate pushes task metrics back to the level of the no-attention baseline, the claim that the scheme preserves task-relevant information under semantic communication would not hold for realistic channels.

Watch

Extended reading notes

Core claim

The central claim is that a coarse semantic condition plus a sparse, attention-masked edge map is enough to regenerate images that satisfy a receiver's current task, and that the selection of which edges to send can be driven entirely by receiver feedback. In the first phase the transmitter sends the segmentation map $x_{seg}$; in the second the receiver's diffusion model produces a coarse reconstruction, and the receiver returns either a class label or a textual prompt; in the third the transmitter computes an attention map $A$ (from CAM or CLIP), forms the masked edge map $x_{att}=M_A\odot x_{edge}$, and transmits only the non-zero $n\times n$ patches of $x_{att}$. The diffusion model then regenerates the image conditioned on both $x_{seg}$ and $x_{att}$. On Cityscapes, this raises detection mIoU for the requested class, lowers object-count MSE, and improves depth RMSE and SI-RMSE relative to no-attention transmission, while compression rates stay in the 19-23 range, between the no-update rate of about 31 and the all-details rate of about 15.

Load-bearing premise

The framework assumes the segmentation map, the masked edge patches, and the receiver feedback all arrive at the other side without bit errors; the experiments contain no wireless channel or noise model, so a corrupted message would degrade the regenerated image and task performance in ways this paper does not measure.

Editorial extensions

If this is right

  • A receiver can switch downstream tasks, for example from car detection to person detection, without requesting a new full image; the transmitter sends only masked edge patches inside the new region of interest and the diffusion model regenerates accordingly.
  • Task-specific updates reduce object-count error sharply: person-count MSE drops from 4.954 with no attention to 1.513 with person-CAM attention, and car detection mIoU rises from 82.37 to 85.21 with car-CAM attention.
  • Compression stays efficient because only non-zero edge patches are transmitted: task-adaptive runs achieve compression rates around 19-23, compared with 31.22 for the no-update baseline and 15.46 for sending all edges.
  • CLIP-based feedback gives better whole-image fidelity, while CAM-based feedback gives more task-oriented detection improvements, so designers can choose the feedback mechanism based on whether perceptual quality or task accuracy matters more.
  • The three-phase updating mechanism can be integrated into other generative semantic communication systems such as GESCO and Diff-GO, making the task-adaptive loop a reusable component.

Reading between the lines

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

  • Inference: because only non-zero edge patches and their locations travel over the link, the same update mechanism could carry other sparse side information, such as depth edges, lane markers, or landmark points, for tasks the paper did not test, like lane detection or pedestrian pose estimation.
  • Inference: the lossless-delivery assumption means the paper's headline result is about semantic selection rather than channel robustness; a natural next experiment is to add channel coding or denoising to the semantic messages and measure how the attention-gated updates behave under bit errors.
  • Inference: CLIP-based attention from free-form text could in principle handle unseen or abstract object categories, but the paper notes that CLIP attention can introduce artifacts, suggesting a calibration or fusion step with CAM attention as a possible improvement.
  • Inference: the threshold $\tau$ on the attention map controls how much detail is transmitted and therefore the compression-performance trade-off; sweeping $\tau$ per task could produce an operating curve similar to rate-distortion curves for semantic communication.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. This paper proposes a three-phase task-adaptive semantic communication framework built on conditional diffusion models. In Step 1, the transmitter sends a compressed segmentation map as a coarse semantic representation; the receiver regenerates the image with a DDPM. In Step 2, the receiver sends task feedback in the form of a class label or a text prompt. In Step 3, the transmitter generates an attention map (via CAM or CLIP), masks an edge map with a threshold, and transmits only nonzero patches of the masked edge map; the receiver then performs a second, task-conditioned regeneration. The framework is evaluated on Cityscapes for reconstruction quality (LPIPS, FID) and downstream tasks (car/person object detection, depth estimation), comparing attention variants against no-attention and all-attention sharing, and against semantic synthesis and semantic communication baselines.

Significance. If the central claim were fully supported, this framework would be a practical contribution to task-oriented semantic communications: it is modular, relies on off-the-shelf components (DDPM, CLIP, ResNet-50 CAM), avoids retraining for new tasks, and provides a clear mechanism for partial semantic updates guided by receiver feedback. The three-phase formulation is intuitive and the exposition is clear. However, as detailed in the major comments, the reported experiments do not yet isolate the effect of task-adaptive selection from the effect of simply transmitting more bits, and the task-level comparison against prior semantic-communication baselines is missing. These issues are fixable within the scope of a revision.

major comments (4)
  1. [Table II / Section III-A3] The reported gains of the attention-masked updates over the 'No-Attn (Step-1 Only)' baseline are confounded with bitrate. No-Attn transmits only the segmentation map (compression rate 31.22), whereas every attention-masked variant transmits additional edge patches (compression rates 18.00 to 22.71), i.e., roughly 1.4 to 2.0 times as many bits. With this design, the improvement in object-count MSE and mIoU could be a monotone consequence of transmitting more information rather than evidence that attention-based selection identifies task-relevant information. Please add a bit-matched control condition that transmits the same number of edge patches without task-adaptive selection (e.g., random patches or a fixed central region) and show that the attention-masked selection outperforms this control at equal bitrate.
  2. [Table II / Section IV-C] The manuscript claims task-level superiority over GESCO and Diff-GO in Section IV-C and the Abstract, but Table II contains no rows for these baselines. The only quantitative comparison to GESCO and Diff-GO appears in Table I, which reports perceptual metrics (LPIPS/FID), not the object-detection and depth metrics that support the central claim. Please include task-metric rows for GESCO and Diff-GO under the same evaluation protocol, or substantially qualify the claims.
  3. [Section IV-A / Table II] The experiments report no error bars, confidence intervals, or significance tests for any metric in Table II, and the number of independent runs is not stated. Several differences that support the adaptive-update claim are small (e.g., depth SI-RMSE 0.2420 vs 0.2628; RMSE 0.0588 vs 0.0609), so without variability estimates the ranking is not statistically supported. Please report mean ± standard deviation over multiple seeds and apply a paired significance test for the main comparisons.
  4. [Section III-A3 / Section IV-A] The two key hyperparameters controlling the adaptive update — the attention threshold τ in the mask definition M_A = I(A > τ) and the edge patch size n×n used in Eq. (4) — are never specified in the experimental setup. Because these directly determine the number of transmitted patches and the achieved compression rate (Table II), the experimental results cannot be reproduced or compared against future work. Please state the chosen values and, ideally, provide a sensitivity analysis.
minor comments (6)
  1. [Fig. 2 caption] The caption for Step 2 says 'the transmitter reconstructs the original media signal,' but the reconstruction is performed at the receiver; please correct this.
  2. [Section III-A1] Please clarify how the one-hot encoded segmentation map is compressed before transmission (e.g., PNG, run-length coding) and whether the reported compression rate in Table II includes this overhead.
  3. [Section III-B1] Please state how the edge condition x_att is injected into the DDPM U-Net (e.g., channel concatenation vs cross-attention), since Eq. (5) only gives the conditioning at the level of the joint likelihood.
  4. [Table I] Several rows report identical LPIPS values (e.g., 0.546 for SPADE, CC-FPSE, and SMIS); please verify these numbers because they may indicate transcription errors from the GESCO source.
  5. [Section IV-A] Please specify the exact Cityscapes split used (the standard train/val/test split) and note that images are resized from the original resolution to 256×512.
  6. [Section IV-C] Please define CR (compression rate) in the text before Table II and state how the 'full data size' is measured (e.g., raw RGB bytes before any coding).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the framework's reconstruction and attention modules are externally grounded, and the reported task gains are experimental outcomes rather than constructed equivalences.

full rationale

The three-phase pipeline is built from independent components: the coarse representation is a segmentation map from an external ResNet; the edge map is Canny-based; attention maps come from CAM (Eq. 1) or pretrained CLIP (Eq. 2); and reconstruction is a standard conditional DDPM with likelihood p_theta(x0:T|xseg,xatt) following Ho et al. [15]. None of these equations is defined in terms of the evaluation metrics (MSE counting, mIoU, RMSE) in Table II, and no fitted parameter is renamed as a prediction. The self-citations to Diff-GO [4] and Diff-GO+ [13] are prior-work/baseline references, not justifications of the central premise; the paper states that 'All the diffusion-based results are re-trained under the same setup for fair comparison,' so the Diff-GO comparison is an external benchmark rather than a load-bearing self-citation chain. The most serious experimental concern—that attention rows transmit more edge bits than the No-Attn control (CR 15.46-22.71 vs. 31.22) and that the feedback label matches the evaluated class—is a potential confound in the experimental design, not a circularity in the derivation, because the task labels and bit rates are inputs to the system rather than outputs derived from the evaluation equations. Therefore, no circular step can be exhibited.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard conditional diffusion models (external DDPM baseline), on pretrained segmentation/edge/CLIP/CAM/YOLO models, and on two unreported hyperparameters (attention threshold tau, patch size n). The largest unstated premise is a noiseless channel; the paper also assumes that receiver feedback can always be expressed as a label or text prompt that maps to a correct spatial attention map. No new entities are introduced.

free parameters (2)
  • Attention mask threshold tau
    Defines M_A = I(A > tau) in Eq. (3); the paper calls it 'predefined' but reports no value or sensitivity analysis, and it controls which edges are transmitted.
  • Edge patch size n x n
    Eq. (4) partitions the masked edge map into n x n patches and discards all-zero patches; n is not reported, and the compression ratio depends on it.
assumptions (4)
  • standard math Conditional DDPM forward/reverse process (Eqs. 5-8) from Ho et al. [15] is a valid generative model for the reconstruction module.
    The paper adopts the standard DDPM formulation without modification; this is accepted prior work, not a new mathematical result.
  • domain assumption Receiver's downstream task can be expressed as a discrete class label or a natural-language prompt that maps to a useful spatial attention map via CAM or CLIP.
    Section III-A2 assumes feedback translates into an attention map; if the prompt is ambiguous or the attention model localizes poorly, the edge update may miss relevant regions.
  • domain assumption Semantic representations (segmentation map, masked edge patches) and feedback are transmitted over a lossless channel with no noise or bit errors.
    No channel model is presented in Section IV; the 'semantic communication' claims therefore apply only to ideal noiseless delivery.
  • domain assumption Pretrained YOLOv8 detections on regenerated images are a valid proxy for task success.
    Object counting and mIoU in Section IV-C are computed from YOLOv8 outputs; detector errors are conflated with reconstruction errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Task-Adaptive Semantic Communications with Controllable Diffusion-based Data Regeneration." pith.science (2026). https://pith.science/paper/SG6C5VGY

@misc{pith2026250507980,
  author       = {Pith},
  title        = {Pith review of: Task-Adaptive Semantic Communications with Controllable Diffusion-based Data Regeneration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SG6C5VGY}},
  note         = {Machine review of arXiv:2505.07980}
}
read the original abstract

Semantic communications represent a new paradigm of next-generation networking that shifts bit-wise data delivery to conveying the semantic meanings for bandwidth efficiency. To effectively accommodate various potential downstream tasks at the receiver side, one should adaptively convey the most critical semantic information. This work presents a novel task-adaptive semantic communication framework based on diffusion models that is capable of dynamically adjusting the semantic message delivery according to various downstream tasks. Specifically, we initialize the transmission of a deep-compressed general semantic representation from the transmitter to enable diffusion-based coarse data reconstruction at the receiver. The receiver identifies the task-specific demands and generates textual prompts as feedback. Integrated with the attention mechanism, the transmitter updates the semantic transmission with more details to better align with the objectives of the intended receivers. Our test results demonstrate the efficacy of the proposed method in adaptively preserving critical task-relevant information for semantic communications while preserving high compression efficiency.

Figures

Figures reproduced from arXiv: 2505.07980 by the authors.

Figure 1
Figure 1. Examples of semantic representation for autonomous driving (from [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Diagram of the proposed three-phase semantic communication: 1) Step 1: the transmitter generates a general and coarse semantic representation to [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Two types of attention schemes: 1) class activation map for classifi [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Structure of the diffusion-based reconstruction module. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Object Detection: Results of Reconstructed Images under Different Conditions: people bounded by blue boxes. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 14 canonical work pages

  1. [1]

    Empowering things with intelligence: A survey of the progress, challenges, and opportunities in artificial intelligence of things,

    J. Zhang and D. Tao, “Empowering things with intelligence: A survey of the progress, challenges, and opportunities in artificial intelligence of things,”IEEE Internet of Things Journal, vol. 8, no. 10, pp. 7789–7817, 2021

    2021

  2. [2]

    Artificial intelligence for vehicle-to-everything: A survey,

    W. Tong, A. Hussain, W. X. Bo, and S. Maharjan, “Artificial intelligence for vehicle-to-everything: A survey,”IEEE Access, vol. 7, pp. 10 823– 10 843, 2019

    work page 2019

  3. [3]

    An introduction to deep learning for the physical layer,

    T. O’Shea and J. Hoydis, “An introduction to deep learning for the physical layer,”IEEE Transactions on Cognitive Communications and Networking, vol. 3, no. 4, pp. 563–575, 2017

    2017

  4. [4]

    Diff- GO: diffusion goal-oriented communications with ultra-high spectrum efficiency,

    A. Wijesinghe, S. Zhang, S. Wanninayaka, W. Wang, and Z. Ding, “Diff- GO: diffusion goal-oriented communications with ultra-high spectrum efficiency,” in2024 IEEE International Conference on Communications Workshops (ICC Workshops), Denver, CO, USA, 2024, pp. 1079–1084

    work page 2024

  5. [5]

    Semantic communications: Overview, open issues, and future research directions,

    X. Luo, H.-H. Chen, and Q. Guo, “Semantic communications: Overview, open issues, and future research directions,”IEEE Wireless Communica- tions, vol. 29, no. 1, pp. 210–219, 2022

    work page 2022

  6. [6]

    Task- oriented communications for 6g: Vision, principles, and technologies,

    Y . Shi, Y . Zhou, D. Wen, Y . Wu, C. Jiang, and K. B. Letaief, “Task- oriented communications for 6g: Vision, principles, and technologies,” IEEE Wireless Communications, vol. 30, no. 3, pp. 78–85, 2023

    2023

  7. [7]

    Deep learning enabled semantic communications with speech recognition and synthesis,

    Z. Weng, Z. Qin, X. Tao, C. Pan, G. Liu, and G. Y . Li, “Deep learning enabled semantic communications with speech recognition and synthesis,”IEEE Transactions on Wireless Communications, vol. 22, no. 9, pp. 6227–6240, 2023

    2023

  8. [8]

    Deep learning enabled seman- tic communication systems,

    H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning enabled seman- tic communication systems,”IEEE Transactions on Signal Processing, vol. 69, pp. 2663–2675, 2021

    2021

Show all 26 references
  1. [9]

    Robust semantic communications with masked vq-vae enabled codebook,

    Q. Hu, G. Zhang, Z. Qin, Y . Cai, G. Yu, and G. Y . Li, “Robust semantic communications with masked vq-vae enabled codebook,”IEEE Transactions on Wireless Communications, vol. 22, no. 12, pp. 8707– 8722, 2023

    2023

  2. [10]

    Generative semantic communication: Diffusion models beyond bit recovery,

    E. Grassucci, S. Barbarossa, and D. Comminiello, “Generative semantic communication: Diffusion models beyond bit recovery,”arXiv preprint arXiv:2306.04321, 2023

    2023 arXiv

  3. [11]

    Wireless end-to-end image transmission system using semantic communications,

    M. U. Lokumarambage, V . S. S. Gowrisetty, H. Rezaei, T. Sivalingam, N. Rajatheva, and A. Fernando, “Wireless end-to-end image transmission system using semantic communications,”IEEE Access, vol. 11, pp. 37 149–37 163, 2023

    2023

  4. [12]

    Token communications: A unified framework for cross-modal context- aware semantic communications,

    L. Qiao, M. B. Mashhadi, Z. Gao, R. Tafazolli, M. Bennis, and D. Niyato, “Token communications: A unified framework for cross-modal context- aware semantic communications,”arXiv preprint arXiv:2502.12096, 2025

    2025 arXiv

  5. [13]

    Diff- go+: An efficient diffusion goal-oriented communication system with local feedback,

    A. Wijesinghe, S. Zhang, S. Wanninayaka, W. Wang, and Z. Ding, “Diff- go+: An efficient diffusion goal-oriented communication system with local feedback,”IEEE Transactions on Wireless Communications, pp. 1–1, 2025

    2025

  6. [14]

    Se- mantic segmentation-based semantic communication system for image transmission,

    J. Wu, C. Wu, Y . Lin, T. Yoshinaga, L. Zhong, X. Chen, and Y . Ji, “Se- mantic segmentation-based semantic communication system for image transmission,”Digital Communications and Networks, vol. 10, no. 3, pp. 519–527, 2024

    2024

  7. [15]

    Denoising diffusion probabilistic mod- els,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic mod- els,” inAdvances in Neural Information Processing Systems, vol. 33, Vancouver, BC, Canada, 2020, pp. 6840–6851

    2020

  8. [16]

    Rethinking atrous convolution for semantic image segmentation,

    L.-C. Chen, G. Papandreou, F. Schroff, and H. Adam, “Rethinking atrous convolution for semantic image segmentation,”arXiv preprint arXiv:1706.05587, 2017

    2017 arXiv

  9. [17]

    An improved canny edge detec- tion algorithm,

    W. Rong, Z. Li, W. Zhang, and L. Sun, “An improved canny edge detec- tion algorithm,” in2014 IEEE International Conference on Mechatronics and Automation, 2014, pp. 577–582

    2014

  10. [18]

    Learning deep features for discriminative localization,

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV , USA, 2016, pp. 2921–2929

    2016

  11. [19]

    Learning transfer- able visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, and Others, “Learning transfer- able visual models from natural language supervision,” inInternational Conference on Machine Learning, Virtual, 2021, pp. 8748–8763

    2021

  12. [20]

    The cityscapes dataset for semantic urban scene understanding,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV , USA, 2016, pp. 3213–3223

    2016

  13. [21]

    Semantic image synthesis with spatially-adaptive normalization,

    T. Park, M.-Y . Liu, T.-C. Wang, and J.-Y . Zhu, “Semantic image synthesis with spatially-adaptive normalization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, 2019, pp. 2337–2346

    2019

  14. [22]

    Learning to predict layout-to-image conditional convolutions for semantic image synthesis,

    X. Liu, G. Yin, J. Shao, X. Wang, and Others, “Learning to predict layout-to-image conditional convolutions for semantic image synthesis,” inAdvances in Neural Information Processing Systems, vol. 32, Vancou- ver, BC, Canada, 2019

    2019

  15. [23]

    Semantically multi-modal image synthesis,

    Z. Zhu, Z. Xu, A. You, and X. Bai, “Semantically multi-modal image synthesis,” inProceedings of the IEEE/CVF CVPR, Seattle, W A, USA, 2020, pp. 5467–5476

    2020

  16. [24]

    You only need adversarial supervision for semantic image synthesis,

    V . Sushko, E. Sch ¨onfeld, D. Zhang, J. Gall, B. Schiele, and A. Khoreva, “You only need adversarial supervision for semantic image synthesis,” arXiv preprint arXiv:2012.04781, 2020

    2012 arXiv

  17. [25]

    Semantic image synthesis via diffusion models,

    W. Wang, J. Bao, W. Zhou, D. Chen, D. Chen, L. Yuan, and H. Li, “Semantic image synthesis via diffusion models,”arXiv preprint arXiv:2207.00050, 2022

    2022

  18. [26]

    Real-time flying object detection with yolov8,

    D. Reis, J. Kupec, J. Hong, and A. Daoudi, “Real-time flying object detection with yolov8,”arXiv preprint arXiv:2305.09972, 2023

    2023 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.