Pith. sign in

REVIEW 5 major objections 4 minor 1 cited by

DT-UFC: Universal Large Model Feature Coding via Peaky-to-Balanced Distribution Transformation

T0 review · 5 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read A universally trained feature codec can compress and reconstruct intermediate representations from different large models and tasks, provided each model's features are first reshaped by a learned peaky-to-balanced distribution transform.

desk verdict A modest but real contribution: Lloyd-Max scalar quantization as a per-model pre-transform enables one codec to serve heterogeneous large-model features, and the empirical evaluation is broad enough to support it. read the letter →

arxiv 2506.16495 v2 pith:6QOP3TKA submitted 2025-06-19 cs.MM cs.CV

classification cs.MMcs.CV
keywords universalfeaturecodingpeaky-to-balanceddistributiontransformationlargemodelscompressioncross-modelgeneralizationrate-accuracyDINOv2LLaMA3
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a single feature codec can serve many large models and tasks if the model-specific feature distributions are first reshaped into a common balanced space. It identifies distributional heterogeneity—peaky LLaMA3 and DINOv2 features versus flatter SD3 features—as the obstacle, and proposes a learned non-uniform scalar transform that maps every feature value into 256 bins of roughly equal mass. Once transformed, features from different models look alike to the codec, so one hybrid-trained codec generalizes across tasks and even architectures. The paper reports rate-accuracy gains over task-specific baselines and shows zero-shot cross-task compression works after alignment.

What carries the argument

The load-bearing object is the learned non-uniform scalar transformation $T: \mathbb{R} \to \{0,1,\dots,255\}$ with inverse $T^{-1}(k) = c_k$. It partitions the original feature-value range into 256 regions and maps each to an integer, choosing the regions and centroids to minimize squared transformation distortion via alternating assignment and update steps, like scalar Lloyd-Max quantization. This reshapes peaky per-model histograms into a near-flat distribution over $\{0,\dots,255\}$, so the codec sees inputs with a consistent range and near-uniform occupancy; the paper claims this improves latent utilization, stabilizes entropy modeling, and aligns distributions across models.

What would settle it

Take DINOv2 features and compute per-channel histograms after applying the paper's transform learned from 10 features; if channels retain visibly different scales or shapes, or if re-learning the transform from 1000 features materially changes rate-accuracy, the single-scalar-transform premise is violated.

Watch

Extended reading notes

Core claim

The central claim is that distributional alignment, rather than a more powerful codec, is the key enabler of universal feature coding. The paper defines universal feature coding as compressing intermediate features from diverse large models with one codec and shows that directly training a codec on heterogeneous features fails because encoders internalize distribution-specific priors. Its proposed solution is the peaky-to-balanced transform $T: \mathbb{R} \to \{0,1,\dots,255\}$, a data-driven partition of the feature value range into 256 intervals whose inverse values are the interval centroids; the transform is fit on only 10 randomly chosen feature vectors by alternating assignment and centroid update. After transformation, the KL divergences between feature distributions from different models drop sharply, and a codec trained on transformed features outperforms task-specifically trained baselines on common-sense reasoning, semantic segmentation, and text-to-image generation, and also compresses ResNet-50 features without retraining.

Load-bearing premise

One scalar transform fit on only ten sampled feature vectors is assumed to capture the distribution of every feature channel and every sample from a model, so that all dimensions land in the same balanced space.

Editorial extensions

If this is right

  • A single hybrid-trained codec can compress features from LLaMA3, DINOv2, and SD3 with better rate-accuracy than codecs trained per task.
  • The transform alone, applied before an untrained baseline codec, improves compressibility and accuracy in the CSR task and causes no degradation in the Seg and TTI tasks.
  • Codecs trained on transformed features generalize to unseen tasks in a zero-shot fashion; for example, models trained on Seg or TTI compress CSR features well after alignment.
  • The approach transfers across architectures: hybrid-trained codecs handle ResNet-50 features with near-lossless classification accuracy without retraining.
  • Codecs trained on transformed features remain effective on imperfect or truncated inputs, indicating they learned general feature structure rather than memorized input patterns.

Reading between the lines

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

  • If the scalar transform generalizes to models beyond the three studied, the same plug-and-play module could be shipped as a per-model header of 256 centroids with one shared codec, a practical deployment pattern for distributed inference.
  • The 256-level target space suggests direct compatibility with existing image codecs and entropy coders; one could test whether the transform also improves coding of natural images whose histograms are strongly skewed.
  • Because the transform is fitted on only 10 features, per-channel or covariate shift is a plausible failure mode; a natural extension the paper does not explore is a per-channel or learned parametric transform.
  • The method's success on ResNet-50 implies the transform may act as a general normalizer for learned representations, potentially useful beyond coding, for example in federated or heterogeneous-device settings where models exchange intermediate tensors.
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

5 major / 4 minor

Summary. The paper addresses universal feature coding for large models, proposing DT-UFC: a per-model scalar transformation learned by Lloyd-Max quantization that maps raw feature values into 0..255, with the goal of making feature distributions more balanced and aligned across models. A single learned codec (Hyperprior or ELIC) is then trained on transformed features and evaluated on common-sense reasoning (LLaMA3), semantic segmentation (DINOv2), and text-to-image synthesis (SD3). Experiments include hybrid and task-specific training, zero-shot cross-task evaluation, a control using transformed features with an unmodified baseline codec, and a cross-architecture test on ResNet-50 features. The paper claims that the transform enables better rate-accuracy and cross-model generalization than task-specific baselines.

Significance. The problem is timely and the proposed solution is refreshingly simple: a plug-and-play scalar transform trained with a well-defined Lloyd-Max objective, plus shared codec training. Strengths include a clean formulation, release of source code, evaluation across three tasks, two codecs, hybrid/task-specific training, and zero-shot cross-task and cross-architecture settings, which makes the central empirical claim partly independent of circularity. The main weaknesses are the gap between the stated 'balanced' design criterion and the actual objective, and the absence of error bars and sensitivity analyses; these need to be addressed before the universality claim is fully supported.

major comments (5)
  1. [Section 4.2, Eq. (3)] The optimization objective in Eq. (3) minimizes reconstruction MSE only; it contains no term that encourages the 256 output bins to be equally occupied, so the 'balanced distribution' criterion stated in Section 4.2 is not actually enforced. For a highly peaky input, Lloyd-Max allocates many bins to high-density regions, and the transformed histogram generally remains far from uniform; Table 1 reports KL divergences between the three transformed distributions but not the divergence from uniformity, and Fig. 2 is illustrative rather than quantitative. Since the universality argument rests on the balanced target space, please either add an explicit balance measure (e.g., entropy of the transformed histogram or KL to uniform) and report it, or modify the transformation to enforce balancing, and rerun the key experiments to confirm the gains remain.
  2. [Section 4.2 and Section 5] The transform is fitted using only 10 randomly selected features, and it is applied element-wise to every scalar of the feature tensor. The paper does not report the size of these 10 features (number of scalar samples), any per-channel distributional analysis, or sensitivity of the downstream results to the particular 10-feature draw. If different channels have very different scales or shapes, a single scalar transform may align pooled marginals while leaving per-channel and spatial statistics misaligned, which could affect the universality claim. Please add: (i) the actual number of scalar samples used in fitting, (ii) a sensitivity study varying the number of features (e.g., 1, 10, 100) and repeated random draws, and (iii) if possible, per-channel KL or histogram comparisons.
  3. [Tables 2-4 and Fig. 5] All rate-accuracy comparisons are reported as single runs without error bars or multiple seeds. Because learned codecs are stochastic and some reported differences are small (e.g., Table 3, TTI rows), the claim that DT-UFC 'consistently outperforms' baselines is not statistically supported. Please retrain with at least 3 seeds and report mean and standard deviation (or show dominant curves), and state the variance of the BPFP and accuracy/FID/mIoU metrics.
  4. [Section 5.5, Table 5] The isolation experiments in Table 5 are not presented at matched rates, which makes the claimed improvement ambiguous. For example, in the CSR rows the transformed-feature entries start at BPFP 2.08 while the baseline entries start at 1.35; in the TTI rows the transformed features use more bitrate than the baseline at similar FID values. Please interpolate both curves to a common set of rates (or equivalent operating points) before claiming that the transform alone improves rate-accuracy.
  5. [Section 5.2, Eq. (6)] The bitrate term BPFP is never defined in the paper, and it is unclear whether it includes the cost of transmitting the transform parameters {c_k} (256 floats per model) or only the quantization indices. Please define BPFP precisely and state whether the transform side information is amortized across the test set; if it is not included, quantify the overhead and justify its omission.
minor comments (4)
  1. [Section 1] The text refers to 'GTP-4'; this should be 'GPT-4'.
  2. [Section 3.2] The KL divergence formula contains a corrupted symbol ('˝') and should be written as a sum or integral.
  3. [Table 1] It would help to add a 'Uniform' column so the reader can see how far the transformed distributions are from the claimed balanced target.
  4. [Figure 3] The caption and axis labels in the supplied version contain stray '/uni...' tokens; please ensure the final PDF renders correctly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the transformation is a fitted preprocessing module, but all universality claims are tested against codecs and features outside its fitting data.

full rationale

The paper's derivation chain does not reduce to its inputs. The core component, T in Section 4.2, is a 1D Lloyd-Max quantizer fitted to minimize the transformation distortion in Eq. (3); it is a preprocessing module, not a predictor of downstream rate-accuracy. The claims that transformed distributions are more aligned are supported empirically by the KL divergences in Table 1, and the KL values are not enforced by Eq. (3), so they are not true by construction. The universality experiments in Sections 5.4 and 5.7 use codecs trained on other tasks or on Transformer features and then evaluate on held-out features from different tasks or architectures, giving independent validation of the codec's generalization. The ResNet-50 experiment fits a transform on 10 CNN features but does not retrain the codec, so the codec itself is tested zero-shot. The only self-citation, reference [19], supplies the dataset and unified test conditions; it is used for evaluation infrastructure rather than as a logical premise, and the paper's central derivation does not depend on an unverified uniqueness theorem or on any prior result by the same authors. The weak assumption that a scalar transform learned from 10 features suffices to align all channels is a legitimate robustness or correctness concern, but it is not a circularity: no fitted parameter is renamed as a prediction, and no equation equates the output to the input. The empirical evaluation against task-specific baselines and external codecs (Hyperprior, ELIC) provides independent content, so no meaningful circular step is present.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities. The only new parameters are the per-model quantization tables, which are fitted.

free parameters (4)
  • Per-model Lloyd-Max centroids c_k (256 values) = Learned on 10 random features per task
    These define the transform T and inverse T^{-1}; they are fit to the feature data of each model, so they are data-dependent parameters the method introduces.
  • Number of quantization bins (256) = 256
    Chosen by hand to match 8-bit image codecs; not derived from data.
  • Number of features for transform fitting = 10
    Randomly selected per task; the paper provides no sensitivity analysis.
  • Target value range [0,255] = 0 to 255
    Fixed design choice to reuse image codec machinery.
assumptions (3)
  • domain assumption All dimensions of a feature vector from a given model share a common scalar value distribution, so a single scalar transform T applies to every element.
    Section 4.2 defines T: R -> {0,...,255} and applies it element-wise without per-dimension fitting; heterogeneous per-dimension scales would break this assumption.
  • domain assumption 10 randomly selected features are representative enough to estimate the transform.
    Stated in Section 4.2; no experiments show stability of the learned transform across different random selections.
  • domain assumption The Lloyd-Max MSE-minimizing quantizer yields a 'balanced' distribution that improves codec generalization.
    The paper claims balancing but optimizes only MSE; the connection between MSE-optimal quantization and codec universality is empirical, not proven. Section 4.2 objective Eq. (3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of DT-UFC: Universal Large Model Feature Coding via Peaky-to-Balanced Distribution Transformation." pith.science (2026). https://pith.science/paper/6QOP3TKA

@misc{pith2026250616495,
  author       = {Pith},
  title        = {Pith review of: DT-UFC: Universal Large Model Feature Coding via Peaky-to-Balanced Distribution Transformation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6QOP3TKA}},
  note         = {Machine review of arXiv:2506.16495}
}
read the original abstract

Like image coding in visual data transmission, feature coding is essential for the distributed deployment of large models by significantly reducing transmission and storage burden. However, prior studies have mostly targeted task- or model-specific scenarios, leaving the challenge of universal feature coding across diverse large models largely unexplored. In this paper, we present the first systematic study on universal feature coding for large models. The key challenge lies in the inherently diverse and distributionally incompatible nature of features extracted from different models. For example, features from DINOv2 exhibit highly peaky, concentrated distributions, while those from Stable Diffusion 3 (SD3) are more dispersed and uniform. This distributional heterogeneity severely hampers both compression efficiency and cross-model generalization. To address this, we propose a learned peaky-to-balanced distribution transformation, which reshapes highly skewed feature distributions into a common, balanced target space. This transformation is non-uniform, data-driven, and plug-and-play, enabling effective alignment of heterogeneous distributions without modifying downstream codecs. With this alignment, a universal codec trained on the balanced target distribution can effectively generalize to features from different models and tasks. We validate our approach on three representative large models (LLaMA3, DINOv2, and SD3) across multiple tasks and modalities. Extensive experiments show that our method achieves notable improvements in both compression efficiency and cross-model generalization over task-specific baselines. All source code has been made available at https://github.com/chansongoal/DT-UFC.

Figures

Figures reproduced from arXiv: 2506.16495 by the authors.

Figure 1
Figure 1. The overall framework of DT-UFC. Large models [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualizations of the original feature distribution (1st column), the nonlinear transformation function (2nd column), [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the latent space. Left: latent repre [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Rate-accuracy performance comparisons between the baseline and the proposed method. The left, middle, and right [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Universality comparison between the baseline and the proposed method. The left, middle, and right figures correspond [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID 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. Compressed Feature Quality Assessment: Dataset and Baselines

    cs.CV 2025-06 conditional novelty 7.0 of 10

    The first compressed feature quality assessment benchmark is released, and three standard similarity metrics are shown to correlate inconsistently with task-level semantic distortion.

Reference graph

Works this paper leans on

81 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

  2. [2]

    Saeed Ranjbar Alvar and Ivan V. Bajić. 2019. Multi-Task Learning with Com- pressible Features for Collaborative Intelligence. In ICIP. 1705–1709. doi:10.1109/ ICIP.2019.8803110

  3. [3]

    Saeed Ranjbar Alvar and Ivan V. Bajić. 2020. Bit Allocation for Multi-Task Collaborative Intelligence. In ICASSP. 4342–4346. doi:10.1109/ICASSP40776.2020. 9054770

  4. [4]

    Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, and Simon Lacoste-Julien

    Devansh Arpit, Stanisław Jastrzundefinedbski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S. Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, and Simon Lacoste-Julien. 2017. A closer look at memorization in deep networks (ICML’17). JMLR, 233–242

  5. [5]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)

  6. [6]

    Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston

    Johannes Ballé, David C. Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. 2018. Variational image compression with a scale hyperprior. ArXiv abs/1802.01436 (2018)

  7. [7]

    Yangang Cai, Peiyin Xing, and Xuesong Gao. 2022. High Efficient 3D Convolu- tion Feature Compression. IEEE Transactions on Circuits and Systems for Video Technology (2022), 1–1. doi:10.1109/TCSVT.2022.3200698

  8. [8]

    Qiaoxi Chen, Changsheng Gao, and Dong Liu. 2024. End-to-End Learned Scalable Multilayer Feature Compression For Machine Vision Tasks. In ICIP. 1781–1787. doi:10.1109/ICIP51287.2024.10647798

Show all 81 references
  1. [9]

    Yi-Hsin Chen, Ying-Chieh Weng, Chia-Hao Kao, Cheng Chien, Wei-Chen Chiu, and Wen-Hsiao Peng. 2023. TransTIC: Transferring transformer-based image compression from human perception to machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision...

  2. [10]

    Zhuo Chen, Kui Fan, Shiqi Wang, Ling-Yu Duan, Weisi Lin, and Alex Kot. 2019. Lossy Intermediate Deep Learning Feature Compression and Evaluation. In Proceedings of the 27th ACM International Conference on Multimedia (MM ’19) . Association for Computing Machinery, New York, NY,...

  3. [11]

    Hyomin Choi and Ivan V. Bajić. 2018. Deep Feature Compression for Collaborative Object Detection. In ICIP. 3743–3747. doi:10.1109/ICIP.2018.8451100

  4. [12]

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. 2024. Scaling rectified flow transformers for high-resolution image synthesis. InICML

  5. [13]

    Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. 2022. Image coding for machines with omnipotent feature learning. In ECCV. Springer, 510–528

  6. [14]

    Changsheng Gao, Yiheng Jiang, Li Li, Dong Liu, and Feng Wu. 2024. DMOFC: Discrimination Metric-Optimized Feature Compression. In PCS. 1–5. doi:10.1109/ PCS60826.2024.10566361

  7. [15]

    Changsheng Gao, Yiheng Jiang, Siqi Wu, Yifan Ma, Li Li, and Dong Liu. 2025. IMOFC: Identity-Level Metric Optimized Feature Compression for Identification Tasks. IEEE Transactions on Circuits and Systems for Video Technology 35, 2 (2025), 1855–1869. doi:10.1109/TCSVT.2024.3467124

  8. [16]

    Changsheng Gao, Li Li, Dong Liu, Zhibo Chen, Weiping Li, and Feng Wu. 2022. Two-Step Fast Mode Decision for Intra Coding of Screen Content. IEEE Transac- tions on Circuits and Systems for Video Technology 32, 8 (2022), 5608–5622

  9. [17]

    Changsheng Gao, Zhuoyuan Li, Li Li, Dong Liu, and Feng Wu. 2024. Rethinking the Joint Optimization in Video Coding for Machines: A Case Study. In DCC. 556–556

  10. [18]

    Changsheng Gao, Dong Liu, Li Li, and Feng Wu. 2023. Towards Task-Generic Image Compression: A Study of Semantics-Oriented Metrics. IEEE Transactions on Multimedia 25 (2023), 721–735. doi:10.1109/TMM.2021.3130754

  11. [19]

    Changsheng Gao, Yifan Ma, Qiaoxi Chen, Yenan Xu, Dong Liu, and Weisi Lin

  12. [20]

    Xuri Ge, Fuhai Chen, Joemon M Jose, Zhilong Ji, Zhongqin Wu, and Xiao Liu. 2021. Structured multi-modal feature embedding and alignment for image-sentence retrieval. In ACM MM. 5185–5193

  13. [21]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al . 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025)

  14. [22]

    Sha Guo, Zhuo Chen, Yang Zhao, Ning Zhang, Xiaotong Li, and Lingyu Duan

  15. [23]

    Robert Henzel, Kiran Misra, and Tianying Ji. 2022. Efficient Feature Compression for the Object Tracking Task. In ICIP. 3505–3509. doi:10.1109/ICIP46576.2022. 9897802

  16. [24]

    Chia-Hao Kao, Cheng Chien, Yu-Jen Tseng, Yi-Hsin Chen, Alessandro Gnutti, Shao-Yuan Lo, Wen-Hsiao Peng, and Riccardo Leonardi. 2024. Bridging com- pressed image latents and multimodal large language models. arXiv preprint arXiv:2407.19651 (2024)

  17. [25]

    Yeongwoong Kim, Hyewon Jeong, Janghyun Yu, Younhee Kim, Jooyoung Lee, Se Yoon Jeong, and Hui Yong Kim. 2023. End-to-End Learnable Multi-Scale Feature Compression for VCM. IEEE Transactions on Circuits and Systems for Video Technology (2023), 1–1. doi:10.1109/TCSVT.2023.3302858

  18. [26]

    JoonHo Lee and Gyemin Lee. 2023. Feature Alignment by Uncertainty and Self- Training for Source-Free Unsupervised Domain Adaptation. Neural Networks 161 (2023), 682–692. doi:10.1016/j.neunet.2023.02.009

  19. [27]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning . PMLR, 19730–19742

  20. [28]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning . PMLR, 12888–12900

  21. [29]

    Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. 2021. Align before fuse: Vision and language repre- sentation learning with momentum distillation. Advances in neural information processing systems 34 (2021), 9694–9705

  22. [30]

    Yao Li, Zhuoyuan Li, Li Li, Dong Liu, and Houqiang Li. 2022. Global homogra- phy motion compensation for versatile video coding. In 2022 IEEE International Conference on Visual Communications and Image Processing (VCIP) . IEEE, 1–5

  23. [31]

    Zhuoyuan Li, Jiacheng Li, Yao Li, Li Li, Dong Liu, and Feng Wu. 2024. In-Loop Filtering via Trained Look-Up Tables. In 2024 IEEE International Conference on Visual Communications and Image Processing (VCIP) . 1–5

  24. [32]

    Zhuoyuan Li, Yao Li, Chuanbo Tang, Li Li, Dong Liu, and Feng Wu. 2024. Uni- formly Accelerated Motion Model for Inter Prediction. In 2024 IEEE International Conference on Visual Communications and Image Processing (VCIP) . 1–5

  25. [33]

    Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. 2024. USTC-TD: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481 (2024)

  26. [34]

    Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. 2024. Object Segmentation-Assisted Inter Prediction for Versatile Video Coding. IEEE Transactions on Broadcasting 70, 4 (2024), 1236–1253

  27. [35]

    Jinming Liu, Ruoyu Feng, Yunpeng Qi, Qiuyu Chen, Zhibo Chen, Wenjun Zeng, and Xin Jin. 2024. Rate-distortion-cognition controllable versatile neural image compression. In European Conference on Computer Vision . Springer, 329–348

  28. [36]

    Kang Liu, Dong Liu, Li Li, Ning Yan, and Houqiang Li. 2021. Semantics-to-Signal Scalable Image Compression with Learned Revertible Representations. Int. J. Comput. Vision 129, 9 (Sept. 2021), 2605–2621. doi:10.1007/s11263-021-01491-7

  29. [37]

    Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. 2015. Learning transferable features with deep adaptation networks. In International conference on machine learning. PMLR, 97–105

  30. [38]

    Yifan Ma, Changsheng Gao, Qiaoxi Chen, Li Li, Dong Liu, and Xiaoyan Sun. 2024. Feature Compression With 3D Sparse Convolution. In VCIP. 1–5

  31. [39]

    Rui Mao, Xinmin Feng, Changsheng Gao, Li Li, Dong Liu, and Xiaoyan Sun. 2024. Perceptual Image Compression With Conditional Diffusion Transformers. In VCIP. 1–5

  32. [40]

    Kiran Misra, Tianying Ji, Andrew Segall, and Frank Bossen. 2022. Video Feature Compression for Machine Tasks. In ICME. 1–6. doi:10.1109/ICME52920.2022. 9859894

  33. [41]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El- Nouby, et al. 2023. DINOv2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

  34. [42]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In ICML. PMLR, 8748–8763

  35. [43]

    Hashimoto, and Percy Liang

    Shiori Sagawa*, Pang Wei Koh*, Tatsunori B. Hashimoto, and Percy Liang. 2020. Distributionally Robust Neural Networks. In ICLR

  36. [44]

    Dmitry Senushkin, Nikolay Patakin, Arseny Kuznetsov, and Anton Konushin

  37. [45]

    Xuelin Shen, Haoqiao Ou, and Wenhan Yang. 2024. Image Coding For Machine Via Analytics-Driven Appearance Redundancy Reduction. In 2024 IEEE International Conference on Image Processing (ICIP) . IEEE, 1883–1889

  38. [46]

    Xuelin Shen, Kangsheng Yin, Xu Wang, Yulin He, Shiqi Wang, and Wenhan Yang

  39. [47]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Independent component alignment for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 20083– 20093

  40. [48]

    Chuanbo Tang, Xihua Sheng, Zhuoyuan Li, Haotian Zhang, Li Li, and Dong Liu

  41. [49]

    Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530 (2024)

  42. [50]

    In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Image coding for analytics via adversarially augmented adaptation. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 3605–3609

  43. [51]

    Baochen Sun and Kate Saenko. 2016. Deep coral: Correlation alignment for deep domain adaptation. In Computer vision–ECCV 2016 workshops: Amsterdam, MM ’25, October 27–31, 2025, Dublin, Ireland Changsheng Gao et al. the Netherlands, October 8-10 and 15-16, 2016, proceedings, pa...

  44. [52]

    Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. 2023. Non-semantics suppressed mask learning for unsupervised video semantic compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 13610– 13622

  45. [53]

    In Proceedings of the AAAI Conference on Artificial Intelligence , Vol

    Offline and online optical flow enhancement for deep video compression. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 5118–5126

  46. [54]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

  47. [55]

    Yuan Tian, Guo Lu, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. 2024. A coding framework and benchmark towards low-bitrate video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 8 (2024), 5852– 5872

  48. [56]

    Yuan Tian, Guo Lu, and Guangtao Zhai. 2024. Free-VSC: Free semantics from visual foundation models for unsupervised video semantic compression. In Euro- pean Conference on Computer Vision . Springer, 163–183

  49. [57]

    Weiqian Wang, Ping An, Xinpeng Huang, Kunqiang Huang, and Chao Yang

  50. [58]

    Yuanyishu Tian, Yao Wan, Lingjuan Lyu, Dezhong Yao, Hai Jin, and Lichao Sun

  51. [59]

    Zixi Wang, Fan Li, Yunfei Zhang, and Yuan Zhang. 2024. Low-Rate Feature Compression for Collaborative Intelligence: Reducing Redundancy in Spatial and Statistical Levels. IEEE Transactions on Multimedia 26 (2024), 2756–2771. doi:10.1109/TMM.2023.3303716

  52. [60]

    April 2023

    WG2. April 2023. Call for Proposals on Feature Compression for Video Coding for Machines. ISO/IEC JTC 1/SC 29/WG 2, N282 (April 2023)

  53. [61]

    Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. 2022. Towards Analysis-Friendly Face Representation With Scalable Feature and Texture Compression. IEEE Transactions on Multimedia 24 (2022), 3169–3181. doi:10.1109/TMM.2021.3094300

  54. [62]

    Shuai Wang, Daoan Zhang, Zipei Yan, Jianguo Zhang, and Rui Li. 2023. Feature alignment and uniformity for test time adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 20050–20060

  55. [63]

    Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. 2024. OpenFedLLM: Training large language models on decentralized private data via federated learning. In ACM SIGKDD. 6137–6147

  56. [64]

    Journal of Visual Communication and Image Representation 95 (2023), 103859

    Intermediate deep feature coding for human–machine vision collaboration. Journal of Visual Communication and Image Representation 95 (2023), 103859

  57. [65]

    Xiaoling Wang, Qi Kang, MengChu Zhou, Siya Yao, and Abdullah Abusorrah

  58. [66]

    Domain adaptation multitask optimization.IEEE Transactions on Cybernetics 53, 7 (2022), 4567–4578

  59. [67]

    Zhicong Zhang, Mengyang Wang, Mengyao Ma, Jiahui Li, and Xiaopeng Fan

  60. [68]

    Lingyu Zhu, Binzhe Li, Riyu Lu, Peilin Chen, Qi Mao, Zhao Wang, Wenhan Yang, and Shiqi Wang. 2024. Learned Image Compression for Both Humans and Machines via Dynamic Adaptation. In 2024 IEEE International Conference on Image Processing (ICIP). 1788–1794

  61. [69]

    Xu Wu, Xianxu Hou, Zhihui Lai, Jie Zhou, Ya-nan Zhang, Witold Pedrycz, and Linlin Shen. 2025. A codebook-driven approach for low-light image enhancement. Engineering Applications of Artificial Intelligence 156 (2025), 111115

  62. [70]

    Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. 2021. SSSIC: Semantics-to-Signal Scalable Image Coding With Learned Structural Representations. IEEE Transactions on Image Processing 30 (2021), 8939–8954. doi:10.1109/TIP.2021.3121131

  63. [72]

    Hao-Wei Yeh, Baoyao Yang, Pong C Yuen, and Tatsuya Harada. 2021. SOFA: Source-data-free feature alignment for unsupervised domain adaptation. In Pro- ceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 474–483

  64. [73]

    Kangsheng Yin, Quan Liu, Xuelin Shen, Yulin He, Wenhan Yang, and Shiqi Wang

  65. [75]

    Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals

  66. [78]

    MSFC: Deep Feature Compression in Multi-Task Network. In ICME. 1–6. doi:10.1109/ICME51207.2021.9428258

  67. [80]

    Yanchen Zuo, Changsheng Gao, Dong Liu, Li Li, Yueyi Zhang, and Xiaoyan Sun

  68. [81]

    IEEE Transactions on Circuits and Systems for Video Technology 34, 3 (2024), 1976–1980

    Learned Rate-Distortion Cost Prediction for Ultrafast Screen Content Intra Coding. IEEE Transactions on Circuits and Systems for Video Technology 34, 3 (2024), 1976–1980

  69. [2021]

    Understanding deep learning (still) requires rethinking generalization. Commun. ACM 64, 3 (Feb. 2021), 107–115. doi:10.1145/3446776

  70. [2022]

    ACM Transactions on Intelligent Systems and Technology 13, 4 (2022), 1–26

    FedBERT: When federated learning meets pre-training. ACM Transactions on Intelligent Systems and Technology 13, 4 (2022), 1–26

  71. [2023]

    In Proceedings of the 31st ACM International Confer- ence on Multimedia (Ottawa ON, Canada) (MM ’23)

    Toward Scalable Image Feature Compression: A Content-Adaptive and Diffusion-Based Approach. In Proceedings of the 31st ACM International Confer- ence on Multimedia (Ottawa ON, Canada) (MM ’23). Association for Computing Machinery, New York, NY, USA, 1431–1442. doi:10.1145/3581...

  72. [2024]

    arXiv preprint arXiv:2412.04307 (2024)

    Feature Coding in the Era of Large Models: Dataset, Test Conditions, and Benchmark. arXiv preprint arXiv:2412.04307 (2024)

  73. [2025]

    In Proceedings of the AAAI Conference on Artificial Intelligence, Vol

    Unified Coding for Both Human Perception and Generalized Machine Ana- lytics with CLIP Supervision. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 9517–9525

Pith tools

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