Pith. sign in

REVIEW 4 major objections 6 minor 30 references

Multi-User Generative Semantic Communication with Intent-Aware Semantic-Splitting Multiple Access

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A semantic-splitting framework with a shared one-hot map and private text prompts, jointly optimized by reinforcement learning, outperforms JPEG and single-modality baselines under equal transmit power.

desk verdict Solid architecture, novel combination, but the PPO reward is never specified—the central joint-optimization claim rests on an unstated implementation detail. read the letter →

arxiv 2507.01333 v1 pith:7UKTPPIG submitted 2025-07-02 cs.NI cs.ITmath.IT

Jiayi Lu , Wanting Yang , Zehui Xiong , Rahim Tafazolli , Tony Q.S. Quek , Mérouane Debbah , Dong In Kim This is my paper
classification cs.NIcs.ITmath.IT
keywords semanticcommunicationgenerativeAIsemantic-splittingmultipleaccessrate-splittingbeamformingreinforcementlearningdiffusionmodelefficiencyscore
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 multi-user generative semantic communication framework, SS-MGSC, in which a base station serves several users with overlapping but distinct information needs by splitting the semantic content into a shared one-hot semantic map and per-user private text prompts. The shared map is broadcast once and the private prompts are sent individually, so redundancy across users is reduced while each user still receives personalized content. On the receiving side, a diffusion model guided by the map and the prompt regenerates a full image for each user. The authors define a semantic efficiency score, built from CLIP and LPIPS, that measures both semantic relevance and perceptual fidelity, and they jointly optimize semantic extraction and beamforming with a reinforcement-learning algorithm. The paper's central claim is that this joint, intent-aware splitting achieves higher semantic efficiency, lower bit error rate, and more robust image generation than transmitting JPEG-compressed images, segmentation maps, or image-only and text-only semantic schemes under the same transmit power.

What carries the argument

The machinery has three pieces. First, semantic-splitting multiple access (SSMA): common semantic information is encoded as a one-hot semantic map broadcast to all users, while private semantic information is encoded as text prompts sent to individual users. Second, the semantic efficiency score (SES), defined as CLIP + (1 − LPIPS), which converts generated-image quality into a scalar objective combining semantic relevance and perceptual fidelity. Third, a proximal policy optimization (PPO) agent whose action space covers the beamforming vectors and the number of common and private semantic units, and whose reward is total SES with penalties for exceeding power or missing per-user SES thresholds. This joint design lets the system trade off the volume of transmitted semantic information against per-symbol transmission quality.

What would settle it

Regenerate the output images several times from the same received one-hot map, text bits, and beamforming solution, and measure the spread in SES; if the regenerated-image variance is as large as the reported gaps between SS-MGSC and its baselines, the ranking would not be reproducible. This would show whether the claimed gains come from the transmission design or from randomness in the diffusion generator.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that semantic information for multiple users can be split into a common one-hot representation and private textual prompts, and that this split can be optimized together with beamforming rather than treating semantic extraction and transmission separately. The framework relies on the semantic efficiency score as the optimization target: a user's SES is CLIP + (1 − LPIPS), so maximizing it simultaneously rewards alignment between the generated image and the intended text and perceptual closeness to the original scene. Under this metric, the jointly optimized SS-MGSC is shown to outperform the JPEG transmission of the original image, the transmission of semantic segmentation maps, and schemes that send only the one-hot map or only the text, at the same maximum transmit power. The results also show that the one-hot map is more robust to noise than a segmentation map, and that the combined common-plus-private scheme yields visibly better generated images at the same power budget.

Load-bearing premise

The optimization assumes the semantic efficiency score, measured after image generation, is a stable and predictable function of the number of semantic units sent and the signal quality, but the paper provides no explicit mapping between them.

Editorial extensions

If this is right

  • Multi-user content dissemination can carry one shared semantic stream plus small personalized streams instead of sending full images to every user, cutting redundancy and transmitted data volume.
  • Because one-hot maps remain robust at low power, the common stream can be served with less transmit power than a JPEG or segmentation-map stream without degrading generated-image quality.
  • Replacing Shannon-capacity constraints with semantic-efficiency constraints in rate-splitting-style optimization changes the resource-allocation outcome, pushing power toward semantic content that matters rather than raw bit reliability.
  • The reported BER-to-SES relationship suggests there is an operating point near BER = 1e-3 below which semantic quality degrades sharply, so system design can target that point instead of chasing near-zero BER.
  • Schemes that send only image-like semantic maps or only text prompts underperform the combined split, indicating that the two modalities are complementary for image regeneration.

Reading between the lines

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

  • Although the paper frames SS-MGSC for vehicular networks, the same semantic-splitting principle applies wherever users share a common scene or context but differ in preferences, such as augmented-reality guidance or metaverse content delivery.
  • The SES objective rewards semantic alignment and perceptual similarity but not the downstream task directly; in safety-critical driving tasks, replacing or augmenting SES with a task-completion reward could change the optimal split between map and text.
  • A direct testable extension would be to run the same framework with randomly chosen beamforming and semantic-unit counts instead of the learned policy; the performance gap between the two would quantify how much of the claimed gain comes from the RL optimization itself.
  • Because CLIP and LPIPS are frozen pre-trained models, the metric may be biased toward their training distributions; evaluating on out-of-distribution scenes or with human ratings could show whether the SES ranking reflects actual user-perceived quality.
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

4 major / 6 minor

Summary. The paper proposes a multi-user downlink generative semantic communication framework called SS-MGSC, aimed at V2X content dissemination. The transmitter splits source information into a common one-hot semantic map broadcast to all users and user-specific private text prompts, then precodes these streams for MIMO transmission. Receivers use a ControlNet-enhanced Stable Diffusion decoder to generate personalized images. The paper defines a semantic efficiency score (SES) combining CLIP and LPIPS, formulates a joint optimization of semantic extraction amounts (nc, np,k) and beamforming to maximize total SES subject to power and per-user SES constraints, and proposes a PPO-based algorithm to solve the problem. Simulation results compare SS-MGSC against variants transmitting only one-hot maps, only text, segmentation maps, and JPEG-compressed images, reporting SES, BER, and generated-image visual quality.

Significance. If the central claim holds, the paper offers a useful architectural idea: separating common and private semantic information and jointly optimizing semantic extraction with beamforming under a semantic objective, rather than a conventional capacity-based one. The one-hot map representation and the use of SES as both optimization objective and evaluation metric are coherent design choices, and the simulation results suggest that the proposed method can outperform simple baselines. The main strengths are the concrete system model, the explicit problem formulation, and an attempt to move beyond bit-level metrics in a multi-user generative setting. However, the significance is currently limited by the absence of a described or validated mechanism for computing the SES reward during RL training, and by the lack of statistical rigor in the experimental comparisons.

major comments (4)
  1. [IV-B and Algorithm 1] The reward r_t in Eq. (14) and Algorithm 1 line 10 require evaluating f_k(nc, np,k, SINRc, SINRp,k), which is defined via CLIP and LPIPS on generated images. The paper never states whether this evaluation runs the full ControlNet-plus-Stable-Diffusion decoder and the CLIP/LPIPS computation inside the PPO training loop. If it does, the claimed training budget of 15,000 episodes on a single RTX 4060 is not credible given the cost of diffusion inference; if it uses a surrogate, a precomputed table, or a reduced evaluation schedule, that mechanism is never described or validated against the true SES. This missing link is load-bearing because the central claim that PPO jointly optimizes semantic extraction and beamforming to maximize SES depends on the reward being an accurate or at least systematically validated proxy for f_k.
  2. [III-A and Eq. (11a)] The objective f_k(nc, np,k, SINRc, SINRp,k) is never concretely specified. Equations (5) and (6) define CLIP and LPIPS on the generated image \hat{C}_k, but the paper does not provide a mapping from the optimization variables (nc, np,k, wc, wk) and the channel to \hat{C}_k, nor any approximation of that mapping. Because the optimization problem in (11) and the MDP reward in (14) both rely on this implicit dependence, the formulation is incomplete without at least an explicit generative model of \hat{C}_k given the transmitted bits and SINR, or a stated surrogate model with fidelity validation.
  3. [V-A and Figs. 7-9] The simulation results are reported as single curves and single image samples, with no error bars, no multiple random seeds, and no confidence intervals. Since CLIP and LPIPS scores are computed on stochastic diffusion outputs, and since the PPO training itself is stochastic, the observed gaps between SS-MGSC and the baselines may not be statistically significant. The paper should report means and standard deviations over several independent runs, or otherwise justify that the reported differences are reproducible.
  4. [V-A] The sentence "we limit the total number of semantic inferences to 20" is ambiguous: it does not state the unit (per episode, per training run, per evaluation, or per PPO step) and is not reflected anywhere in Algorithm 1. This ambiguity directly affects the feasibility of the reward computation described in Eq. (14), and it should be clarified with a precise budget and how that budget is enforced.
minor comments (6)
  1. [II-D] There is an unresolved citation placeholder in the sentence "each vehicle has sufficient computational capacity [ ?]"; this reference must be supplied or removed.
  2. [III-B] The statement "it is evident that the problem is NP-hard" is asserted without proof or citation. Either provide a reduction or reference, or soften the claim to say the problem is challenging due to mixed-integer and nonconvex structure.
  3. [IV-A, Eq. (14)] The signs of the reward weights α and β are not specified. If α and β are positive, the penalty terms are non-positive as intended, but the paper should state this explicitly and report their values in Table I.
  4. [II-B and Eqs. (11d)-(11e)] The notation M and N is used both as the set of possible values and as the maximum value in the constraints; use e.g. M_max and N_max, or calligraphic sets, to avoid ambiguity.
  5. [Figs. 5-12] Several figure captions and axis labels appear as unreadable escaped strings (e.g., sequences like "/uni00000013/...") in the manuscript source; the figures should be regenerated with clearly legible text.
  6. [V-A] The text states the transmitted image is "approximately 0.88 Mb or 416 KB," but these two values are inconsistent and should be corrected.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the SES objective is an intentionally chosen metric, not a prediction derived from its own definition; self-citations are present but not load-bearing.

full rationale

The paper's central derivation chain is not circular. The SES metric is defined via CLIP and LPIPS, and then it is used as both the RL reward and the evaluation metric; this is a consistent objective choice, not a case where a predicted quantity reduces to a fitted input by construction. The comparative claims that SS-MGSC outperforms JPEG, SegS-MGSC, O-MGSC, and T-MGSC are empirical results obtained under the same transmit power, not consequences of the metric definition. The RL formulation maximizes the reward in Eq. (14), which is the SES objective plus penalty terms, while the evaluation in Figs. 7-9 reports the same SES; this alignment is expected for any optimization work and does not make the performance comparison tautological, since PPO convergence to a global optimum is not guaranteed and the baselines are not optimized with the same policy. The paper does not fit parameters to data and then rename that fit as a prediction. One genuine weakness is that the reward computation during training is underspecified: Section V-A limits the total number of semantic inferences to 20, but Algorithm 1 line 10 only says 'calculate the reward rt' without explaining how SES is evaluated inside the loop. That is an implementation and reproducibility gap, not circularity. Self-citations, including the reference to the authors' earlier work [16] and co-authored surveys [1], [3], are present, but they are used for motivation and background, not as load-bearing evidence for the framework's claims. The framework is self-contained against external datasets and baseline schemes, so the appropriate circularity finding is low.

Assumptions & free parameters 2 free parameters · 3 assumptions · 2 invented entities

The framework rests on a small set of domain assumptions about onboard compute, the SKB, and the validity of the SES metric. No new physical entities are proposed. The main free parameters are the reward weighting factors and SES thresholds, both unreported.

free parameters (2)
  • alpha, beta (reward penalty weights) = not reported
    Eq. (14) in Section IV.A.4: the reward function couples the objective and constraint penalties with weighting factors alpha and beta, whose values are never given in the paper or in Table I.
  • SES threshold I_th,k = not specified
    Constraint (11c) and the reward penalty in (14) require per-user SES thresholds, but the paper never states how these thresholds are chosen or whether they are achievable.
assumptions (3)
  • domain assumption Users have sufficient computational capacity to run ControlNet-enhanced diffusion
    Section II-D states 'each vehicle has sufficient computational capacity [ ?] and is equipped with a generative model'. The missing citation and unstated hardware requirements are a load-bearing assumption for receiver-side generation.
  • domain assumption The intent-aware SKB accurately represents user preferences and intents, and the wireless channel can be modeled by Rayleigh fading with known large-scale path loss
    Section II-A and II-C. The entire common/private split relies on the SKB having correct prior knowledge; the channel model is standard, but the paper provides no validation against real channel traces.
  • ad hoc to paper SES is a monotonic proxy for communication quality such that maximizing CLIP + (1-LPIPS) improves user satisfaction
    Section IV-A defines SES as this linear combination. There is no external validation or ablation showing that this metric correlates with task performance or user preference.
invented entities (2)
  • Intent-aware shared knowledge base (SKB)
    purpose: Provides prior knowledge of user preferences and intents to guide semantic encoding and decoding in the SS-MGSC framework.
    Section II-A and Fig. 3 describe its construction using LLMs and knowledge graphs, but no dataset or experiment validates its accuracy or its necessity; it is a postulated system component.
  • Semantic efficiency score (SES)
    purpose: Optimization objective and evaluation metric combining CLIP and LPIPS.
    Defined in Section IV-A as CLIP + (1-LPIPS). It is a new composite metric with no external validation that higher SES corresponds to better communication or task performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-User Generative Semantic Communication with Intent-Aware Semantic-Splitting Multiple Access." pith.science (2026). https://pith.science/paper/7UKTPPIG

@misc{pith2026250701333,
  author       = {Pith},
  title        = {Pith review of: Multi-User Generative Semantic Communication with Intent-Aware Semantic-Splitting Multiple Access},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7UKTPPIG}},
  note         = {Machine review of arXiv:2507.01333}
}
read the original abstract

With the booming development of generative artificial intelligence (GAI), semantic communication (SemCom) has emerged as a new paradigm for reliable and efficient communication. This paper considers a multi-user downlink SemCom system, using vehicular networks as the representative scenario for multi-user content dissemination. To address diverse yet overlapping user demands, we propose a multi-user Generative SemCom-enhanced intent-aware semantic-splitting multiple access (SS-MGSC) framework. In the framework, we construct an intent-aware shared knowledge base (SKB) that incorporates prior knowledge of semantic information (SI) and user-specific preferences. Then, we designate the common SI as a one-hot semantic map that is broadcast to all users, while the private SI is delivered as personalized text for each user. On the receiver side, a diffusion model enhanced with ControlNet is adopted to generate high-quality personalized images. To capture both semantic relevance and perceptual similarity, we design a novel semantic efficiency score (SES) metric as the optimization objective. Building on this, we formulate a joint optimization problem for multi-user semantic extraction and beamforming, solved using a reinforcement learning-based algorithm due to its robustness in high-dimensional settings. Simulation results demonstrate the effectiveness of the proposed scheme.

Figures

Figures reproduced from arXiv: 2507.01333 by the authors.

Figure 1
Figure 1. System model of multi-user Generative SemCom in vehicular network. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The detailed architecture of the Intent-aware SS-MGSC. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Construction of Intent-aware SKB. [w1, w2, . . . , wK]. The transmitted signal from the transmitter can be expressed as x = Ws. (2) For the kth user, the received signal can be represented as yk = h H BU,kx + zk = h H BU,kwksk + X K l̸=k h H BU,kwlsl + zk, (3) where zk ∼ CN 0, σ2  denotes the additive white Gaussian noise (AWGN) received by kth user and the second term in (3) represents the multi-user interference … view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: The PPO-based SS-MGSC framework. aim of reinforcement learning is to maximize the long-term accumulated reward, which is defined as: R = X∞ l=0 γ k rt+l , (15) where l represents the time step offset from the current time step t. 5) Policy: The policy πθ (st, at) speci…
Figure 5
Figure 5. Figure 5: Convergence performance of the PPO-based SES-oriented SSMA [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Convergence performance of the PPO-based SES-oriented SSMA [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Performance comparison of different schemes for PPO-based SES [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 9
Figure 9. Figure 9: Generated images under different maximum transmit power levels for SS-MGSC, SegS-MGSC, and JPEG transmission. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: BER comparison of image and text transmission in the SS-MGSC [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: SES comparison for different users under image BER in the SS [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: SES comparison for different users under text BER in the SS-MGSC [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 26 canonical work pages

  1. [1]

    Semantic communications for future internet: Fundamen- tals, applications, and challenges,

    W. Yang et al., “Semantic communications for future internet: Fundamen- tals, applications, and challenges,” IEEE Commun. Surveys Tuts., vol. 25, no. 1, pp. 213–250, Jan. 2022

    work page 2022

  2. [2]

    Deep learning enabled semantic communication systems,

    H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Trans. Signal Process., vol. 69, pp. 2663–2675, Apr. Apr. 2021

    work page 2021

  3. [3]

    Rethinking generative semantic communication for multi- user systems with large language models,

    W. Yang et al., “Rethinking generative semantic communication for multi- user systems with large language models,” IEEE Wireless Commun., pp. 1–9, Apr. 2025. 12

    work page 2025

  4. [4]

    Scsc: A novel standards-compatible semantic communica- tion framework for image transmission,

    X. Han et al., “Scsc: A novel standards-compatible semantic communica- tion framework for image transmission,” IEEE Trans. Commun., pp. 1–1, Jan. 2025

    work page 2025

  5. [5]

    Generative ai for secure physical layer communications: A survey,

    C. Zhao et al., “Generative ai for secure physical layer communications: A survey,” IEEE Trans. Cogn. Commun. Netw. , Aug. 2024

    work page 2024

  6. [6]

    Image segmentation semantic communication over internet of vehicles,

    Q. Pan et al., “Image segmentation semantic communication over internet of vehicles,” in in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC) , Jul. 2023, pp. 1–6

    work page 2023

  7. [7]

    Semantic communication systems for speech transmission,

    Z. Weng and Z. Qin, “Semantic communication systems for speech transmission,” IEEE J. Sel. Areas Commun. , vol. 39, no. 8, pp. 2434– 2444, Jun. Jun. 2021

    work page 2021

  8. [8]

    Evolving semantic communication with generative modelling,

    S. Tang, Q. Yang, D. G ¨und¨uz, and Z. Zhang, “Evolving semantic communication with generative modelling,” in in Proc. IEEE 35th Int. Symp. Personal, Indoor Mobile Radio Commun. (PIMRC), Jan. Jan. 2024, pp. 1–6

    work page 2024

Show all 30 references
  1. [9]

    Deep joint semantic coding and beamforming for near- space airship-borne massive MIMO network,

    M. Wu et al. , “Deep joint semantic coding and beamforming for near- space airship-borne massive MIMO network,” 2024, arXiv:2405.19889

    2024 arXiv

  2. [10]

    Energy efficient semantic communication over wireless networks with rate splitting,

    Z. Yang, M. Chen, Z. Zhang, and C. Huang, “Energy efficient semantic communication over wireless networks with rate splitting,” IEEE J. Sel. Areas Commun., vol. 41, no. 5, pp. 1484–1495, Jan. 2023

    2023

  3. [11]

    Scan: Semantic communication with adaptive channel feedback,

    G. Zhang, Q. Hu, Y . Cai, and G. Yu, “Scan: Semantic communication with adaptive channel feedback,” IEEE Trans. Cogn. Commun. Netw. , Apr. 2024

    2024

  4. [12]

    Deep learning-empowered semantic communication systems with a shared knowledge base,

    P. Yi, Y . Cao, X. Kang, and Y .-C. Liang, “Deep learning-empowered semantic communication systems with a shared knowledge base,” IEEE Trans. Wireless Commun., vol. 23, no. 6, pp. 6174–6187, Nov. 2023

    2023

  5. [13]

    Bert: Pre- training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre- training of deep bidirectional transformers for language understanding,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolog...

    2019

  6. [14]

    Gpt-4 technical report,

    J. Achiam et al., “Gpt-4 technical report,” 2023, arXiv:2303.08774

    2023 arXiv

  7. [15]

    Rate-splitting multiple access: Fundamentals, survey, and future research trends,

    Y . Mao et al., “Rate-splitting multiple access: Fundamentals, survey, and future research trends,” IEEE Commun. Surv. Tutor. , vol. 24, no. 4, pp. 2073–2126, July. 2022

    2022

  8. [16]

    Generative artificial intelligence-enhanced multimodal semantic communication in internet of vehicles: System design and methodologies,

    J. Lu et al. , “Generative artificial intelligence-enhanced multimodal semantic communication in internet of vehicles: System design and methodologies,” IEEE Veh. Technol. Mag. , vol. 20, no. 2, pp. 71–82, Mar. 2025

    2025

  9. [17]

    Select2col: Leveraging spatial-temporal importance of semantic information for efficient collaborative perception,

    Y . Liu et al. , “Select2col: Leveraging spatial-temporal importance of semantic information for efficient collaborative perception,” IEEE Veh. Technol., Apr. 2024

    2024

  10. [18]

    Resource allocation and common message selection for task-oriented semantic information transmission with rsma,

    Y . Cheng et al., “Resource allocation and common message selection for task-oriented semantic information transmission with rsma,” IEEE Trans. Wireless Commun., Oct. 2023

    2023

  11. [19]

    A survey of large language models,

    W. X. Zhao et al. , “A survey of large language models,” 2023, arXiv:2303.18223

    2023 arXiv

  12. [20]

    Unified perceptual pars- ing for scene understanding,

    T. Xiao, Y . Liu, B. Zhou, Y . Jiang, and J. Sun, “Unified perceptual pars- ing for scene understanding,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 418–434

    2018

  13. [21]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie et al. , “Segformer: Simple and efficient design for semantic segmentation with transformers,” Adv. Neural Inf. Process. Syst., vol. 34, pp. 12 077–12 090, 2021

    2021

  14. [22]

    Prompting hard or hardly prompting: Prompt inversion for text-to-image diffusion models,

    S. Mahajan, T. Rahman, K. M. Yi, and L. Sigal, “Prompting hard or hardly prompting: Prompt inversion for text-to-image diffusion models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , Jun. 2024, pp. 6808–6817

    2024

  15. [23]

    Hard prompts made easy: Gradient-based discrete opti- mization for prompt tuning and discovery,

    Y . Wen et al., “Hard prompts made easy: Gradient-based discrete opti- mization for prompt tuning and discovery,” in Adv. Neural Inf. Process. Syst., vol. 36, Dec. 2023, pp. 51 008–51 025

    2023

  16. [24]

    Recti- fied diffusion: Straightness is not your need in rectified flow,

    F.-Y . Wang, L. Yang, Z. Huang, M. Wang, and H. Li, “Recti- fied diffusion: Straightness is not your need in rectified flow,” 2024, arXiv:2410.07303

    2024 arXiv

  17. [25]

    Instaflow: One step is enough for high-quality diffusion- based text-to-image generation,

    X. Liu et al., “Instaflow: One step is enough for high-quality diffusion- based text-to-image generation,” in in Proc. 12th Int. Conf. Learn. Represent. (ICLR)., May. 2023

    2023

  18. [26]

    Learning transferable visual models from natural language supervision,

    A. Radford et al. , “Learning transferable visual models from natural language supervision,” in in Proc. 38th Int. Conf. Mach. Learn. (ICML) , May. 2021, pp. 8748–8763

    2021

  19. [27]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang et al. , “The unreasonable effectiveness of deep features as a perceptual metric,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2018, pp. 586–595

    2018

  20. [28]

    Image quality metrics: PSNR vs. SSIM,

    A. Hore and D. Ziou, “Image quality metrics: PSNR vs. SSIM,” in Proc. 20th Int. Conf. Pattern Recognit. (ICPR) , Aug. 2010, pp. 2366–2369

    2010

  21. [29]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” 2017, arXiv:1707.06347

    2017 arXiv

  22. [30]

    Rain removal in traffic surveillance: Does it matter?

    C. H. Bahnsen and T. B. Moeslund, “Rain removal in traffic surveillance: Does it matter?” IEEE Trans. Intell. Transp. Syst. , vol. 20, no. 8, pp. 2802–2819, Oct. 2018

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.