Pith. sign in

REVIEW 3 major objections 2 minor 50 references

Grouped Speculative Decoding for Autoregressive Image Generation

T0 review · 3 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Grouped Speculative Decoding accelerates autoregressive image models by 3.7x on average with no quality loss and no additional training.

desk verdict Plausible cluster-acceptance idea, but unverifiable as supplied — the full text is a different paper, so the 3.7x speedup claim is hanging without experiments. read the letter →

arxiv 2508.07747 v1 pith:IRZZMG2B submitted 2025-08-11 cs.CV

Junhyuk So , Juncheol Shin , Hyunho Kook , Eunhyeok Park This is my paper
classification cs.CV
keywords speculativedecodingautoregressiveimagegenerationtraining-freeaccelerationtokenclusteringredundancygroupedinferencespeedup
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 claims that the standard speculative decoding recipe, which accepts or rejects a single most-likely token at each step, is a poor fit for images because image tokens carry redundancy: several different tokens can express visually valid content at the same position. The authors introduce Grouped Speculative Decoding (GSD), which instead accepts clusters of visually valid tokens, cutting the false-negative rejections that slow down generation. They report an average 3.7x wall-clock speedup across autoregressive image models, with image quality preserved, and no training required. If true, this gives a practical, drop-in acceleration for sequential image generation.

What carries the argument

The load-bearing mechanism is a cluster-based acceptance rule for speculative decoding, applied to image-token sequences. Instead of checking whether the draft token equals the argmax token of the target model, GSD first builds clusters of tokens that are mutually valid visual alternatives at the current step, then accepts the draft token if the target model's predicted token falls in the same cluster. The dynamic clustering (as opposed to a static embedding-distance clustering, which the paper shows to be ineffective) is what keeps the acceptance rate high without sacrificing quality.

What would settle it

Run GSD on a high-detail image dataset (e.g., faces or text-heavy images) and compare FID or CLIP score against the original autoregressive model as the cluster size grows. If generated-image quality degrades monotonically with cluster size, the redundancy premise is false; if quality stays flat, it is confirmed. A more direct test: measure the per-step KL divergence between the GSD-modified token distribution and the target model's distribution across a long sequence; a per-step divergence that compounds into visible distribution shift would falsify the claim.

Watch

Extended reading notes

Core claim

GSD rests on the observation that image tokens are not like language tokens: at a given position, multiple token choices can carry equivalent visual semantics. Traditional speculative decoding verifies the draft model's prediction against the single most-likely token of the target model, so any alternative token that is visually plausible but not the argmax is treated as an error, producing excessive rejections. GSD replaces that rule with a cluster-based rule: it forms groups of tokens that are visually valid substitutions, and accepts a draft prediction if the target model's distribution falls inside the same cluster. The paper further finds that static clustering by embedding distance is

Load-bearing premise

The claim that quality is preserved rests on the assumption that image-token positions really do have multiple tokens that carry equivalent visual meaning, so swapping in a cluster member instead of the argmax token never measurably changes the final image.

Editorial extensions

If this is right

  • Any autoregressive image generation model can be accelerated with GSD without retraining, as long as it provides both a draft and target model.
  • The average speedup of 3.7x is a baseline; models with more redundant token distributions will see larger gains.
  • Static clustering based on embedding distance is not sufficient; the dynamic scheme is necessary to preserve quality.
  • Quality is preserved because accepted cluster members are visually equivalent tokens, so the output distribution shifts only within perceptually valid alternatives.

Reading between the lines

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

  • A natural extension the paper does not pursue is to treat cluster size as a tunable knob and quantify the quality-speed trade-off curve, rather than the single operating point reported.
  • The underlying diagnosis that single-token acceptance over-rejects in redundant token distributions likely transfers to other sequential generative models over visual or audio tokens, making GSD part of a broader principle rather than an image-specific trick.
  • A testable prediction follows from the redundancy premise: models trained on low-entropy, repetitive image distributions should enjoy larger speedups under GSD than models on high-detail distributions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 2 minor

Summary. The abstract describes Grouped Speculative Decoding (GSD), a training-free method to accelerate autoregressive image generation by accepting clusters of visually valid tokens during speculative decoding, claiming an average 3.7x speedup while preserving image quality. However, the full text supplied with this submission is not the GSD paper at all: it is a different manuscript, 'Correspondence as Video: Test-Time Adaption on SAM2 for Reference Segmentation in the Wild' (arXiv:2508.07759), with a different title, different authors, and entirely different method and experiments. Consequently, the submitted manuscript contains no description of the GSD algorithm, no derivation of its properties, no experimental protocol, and no results supporting the 3.7x speedup or the quality-preservation claim.

Significance. If the claims in the abstract were substantiated, the idea of accepting clusters of visually valid tokens rather than a single most-likely token would be an interesting and potentially impactful extension of speculative decoding to image generation, especially because it is claimed to be training-free and to give a large speedup. The conceptual move of leveraging image-token redundancy is plausible. However, the submitted manuscript provides no evidence for these claims: the only substantive content is an unrelated reference-segmentation paper. There are no derivations, no machine-checked proofs, no experimental tables, and no reproducibility artifacts for GSD in this submission. The significance of the claimed result cannot be assessed from the provided materials.

major comments (3)
  1. [Full Text (all sections)] The full text attached to this submission is arXiv:2508.07759, a paper on reference segmentation with SAM2 titled 'Correspondence as Video: Test-Time Adaption on SAM2 for Reference Segmentation in the Wild.' It contains no mention of Grouped Speculative Decoding, no autoregressive image generation, and no speculative decoding. The central claim of the abstract (an average 3.7x speedup with preserved quality, training-free) is therefore wholly unsupported by the submitted manuscript. This is not a local omission that can be repaired by editing a section; the substance of the paper is absent.
  2. [Abstract, quality-preservation claim] The abstract asserts that 'image tokens exhibit inherent redundancy and diversity' and that evaluating 'clusters of visually valid tokens' avoids 'excessive false-negative rejections' while 'preserving image quality.' However, a cluster-based acceptance rule that accepts a non-argmax token from an accepted cluster changes the sampling distribution at every accepted position, whereas standard speculative decoding preserves the target distribution by falling back to the target distribution upon rejection. No analysis is provided, in the abstract or elsewhere, bounding the divergence between the GSD-induced distribution and the target distribution over a full autoregressive sequence. Even a plausible per-step redundancy argument does not by itself establish global distribution preservation, so the quality-preservation claim is load-bearing and unsubstantiated.
  3. [Abstract, 3.7x speedup] The headline number, 'accelerates AR image models by an average of 3.7x,' is an empirical measurement. The submitted full text contains no experimental section, no datasets, no baselines, no hardware description, and no measurement of the dynamic-clustering overhead. Without the experimental protocol and results, the 3.7x figure cannot be verified, reproduced, or even meaningfully compared with existing speculative decoding baselines. This is a load-bearing evidentiary gap, not a presentation issue.
minor comments (2)
  1. [Title/Abstract vs. Full Text] The title and author list of the supplied full text do not match the title and abstract of the submission. The abstract advertises GSD and cites a GitHub repository, but the full text is a different paper with its own abstract, contributions, and references. This mismatch makes the submission internally inconsistent and unreviewable as a coherent manuscript.
  2. [References] The reference list in the full text pertains entirely to reference segmentation and SAM2, with no references to speculative decoding or autoregressive image generation. If the correct GSD paper is subsequently provided, the references will need to be completely redone.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; supplied full text is a different paper (arXiv:2508.07759) so no derivation chain for GSD is present to exhibit reduction.

full rationale

The abstract of arXiv:2508.07747 claims a training-free 3.7x average speedup via Grouped Speculative Decoding, with the central premise being that image tokens exhibit redundancy and diversity and that cluster-based acceptance avoids false-negative rejections. The supplied full text, however, is arXiv:2508.07759, 'Correspondence as Video: Test-Time Adaption on SAM2 for Reference Segmentation in the Wild,' which contains no GSD method, no equations for cluster acceptance, no acceptance-rejection analysis, and no experiments measuring wall-clock speedup or image quality. Consequently, there is no derivation chain to walk: the abstract's 3.7x figure is presented as an empirical measurement, not as a quantity derived from fitted parameters or from a self-citation. No specific step can be quoted that reduces to its own inputs. The absence of method and experiments is an evidentiary gap (correctness risk), not circularity. Under the hard rules, circularity may only be claimed with a quoted reduction or fitted-parameter-as-prediction; none exists here. Therefore the honest finding is no significant circularity, score 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

This ledger is provisional because the provided full text is a different paper (CAV-SAM, arXiv:2508.07759); the GSD manuscript could not be read. The free parameters and axioms below are reconstructed from the abstract's description of GSD: the method requires hyperparameters for cluster formation and acceptance, and the speedup number depends on the draft/target model pairing. The two domain axioms, token redundancy and quality-preserving cluster acceptance, are the load-bearing premises of the method and are asserted in the abstract without supporting analysis in the available material. No invented entities beyond a cluster-decision rule are visible at this level.

free parameters (2)
  • Cluster definition and acceptance hyperparameters (cluster size, distance threshold, dynamic-interpolation schedule) = not reported in abstract
    GSD must decide which tokens form a 'visually valid' cluster and when a cluster is accepted. These choices trade off speed (larger clusters accept more) against quality (over-acceptance drifts the distribution). The abstract reports no values.
  • Draft model and verification batch size = not reported in abstract
    The 3.7x figure is an end-to-end speedup that depends on draft-vs-target model cost and the number of tokens verified per step, neither of which is stated in the abstract.
assumptions (3)
  • domain assumption Image tokens are multiply-satisfiable: for a given context many different next tokens are semantically valid, so rejecting all but the argmax token creates false negatives.
    Stated in the abstract ('multiple tokens can convey valid semantics'); it is the reason cluster acceptance can be quality-preserving. If it fails, GSD has no advantage over standard SD.
  • domain assumption Accepting an alternative token inside an approved cluster preserves the output image distribution well enough to claim 'preserving image quality' without training.
    The abstract's quality-preservation claim requires that cluster-level acceptance does not introduce drift; no distributional argument or measurement appears in the abstract.
  • domain assumption Speculative decoding's rejection-sampling correctness carries over to the cluster setting (the draft's proposals are the only sampled distribution and the target's cluster-level verification maintains the target distribution).
    GSD is built on prior SD work that the abstract cites but does not restate; the formal acceptance condition for clusters is not given in the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Grouped Speculative Decoding for Autoregressive Image Generation." pith.science (2026). https://pith.science/paper/IRZZMG2B

@misc{pith2026250807747,
  author       = {Pith},
  title        = {Pith review of: Grouped Speculative Decoding for Autoregressive Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IRZZMG2B}},
  note         = {Machine review of arXiv:2508.07747}
}
read the original abstract

Recently, autoregressive (AR) image models have demonstrated remarkable generative capabilities, positioning themselves as a compelling alternative to diffusion models. However, their sequential nature leads to long inference times, limiting their practical scalability. In this work, we introduce Grouped Speculative Decoding (GSD), a novel, training-free acceleration method for AR image models. While recent studies have explored Speculative Decoding (SD) as a means to speed up AR image generation, existing approaches either provide only modest acceleration or require additional training. Our in-depth analysis reveals a fundamental difference between language and image tokens: image tokens exhibit inherent redundancy and diversity, meaning multiple tokens can convey valid semantics. However, traditional SD methods are designed to accept only a single most-likely token, which fails to leverage this difference, leading to excessive false-negative rejections. To address this, we propose a new SD strategy that evaluates clusters of visually valid tokens rather than relying on a single target token. Additionally, we observe that static clustering based on embedding distance is ineffective, which motivates our dynamic GSD approach. Extensive experiments show that GSD accelerates AR image models by an average of 3.7x while preserving image quality-all without requiring any additional training. The source code is available at https://github.com/junhyukso/GSD

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 37 canonical work pages

  1. [1]

    Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot seg- mentation without meta-learning: A good transductive infer- ence is all you need? In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 13979–13988, 2021. 7

    work page 2021

  2. [2]

    Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration

    Sema Candemir, Stefan Jaeger, Kannappan Palaniappan, Jonathan P Musco, Rahul K Singh, Zhiyun Xue, Alexandros Karargyris, Sameer Antani, George Thoma, and Clement J McDonald. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. IEEE Transac- tions on Medical Imaging, 33(2):577–590, 2013. 6

    work page 2013

  3. [3]

    Pixel matching network for cross-domain few- shot segmentation

    Hao Chen, Yonghan Dong, Zheming Lu, Yunlong Yu, and Jungong Han. Pixel matching network for cross-domain few- shot segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 978– 987, 2024. 7

    work page 2024

  4. [4]

    Holistic pro- totype activation for few-shot segmentation

    Gong Cheng, Chunbo Lang, and Junwei Han. Holistic pro- totype activation for few-shot segmentation. IEEE Transac- tions on Pattern Analysis and Machine Intelligence , 45(4): 4650–4666, 2022. 1, 3

    work page 2022

  5. [5]

    Tracking anything with decoupled video segmentation

    Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexan- der Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 1316–1326, 2023. 2, 8

    work page 2023

  6. [6]

    Segment and track anything

    Yangming Cheng, Liulei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 2

    arXiv 2023

  7. [7]

    Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the interna- tional skin imaging collaboration (isic)

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the interna- tional skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 6

    arXiv 2018

  8. [8]

    Deepglobe 2018: A challenge to parse the earth through satellite images

    Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. Deepglobe 2018: A challenge to parse the earth through satellite images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition workshops, pages 172–181, 2018. 6

    work page 2018

Show all 50 references
  1. [9]

    Self- support few-shot semantic segmentation

    Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self- support few-shot semantic segmentation. In European Con- ference on Computer Vision, pages 701–719. Springer, 2022. 7

    2022

  2. [10]

    Apseg: Auto-prompt network for cross-domain few-shot semantic segmentation

    Weizhao He, Yang Zhang, Wei Zhuo, Linlin Shen, Jiaqi Yang, Songhe Deng, and Liang Sun. Apseg: Auto-prompt network for cross-domain few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23762–23772, 2024. 1, 2, 3, 7

    2024

  3. [11]

    Adapt before comparison: A new perspective on cross-domain few-shot segmentation

    Jonas Herzog. Adapt before comparison: A new perspective on cross-domain few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23605–23615, 2024. 1, 3, 7

    2024

  4. [12]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. Advances in Neural Information Processing Systems, 33:6840–6851, 2020. 3

    2020

  5. [13]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3

    2021 arXiv

  6. [14]

    Tfmq-dm: Temporal feature maintenance quantization for diffusion models

    Yushi Huang, Ruihao Gong, Jing Liu, Tianlong Chen, and Xianglong Liu. Tfmq-dm: Temporal feature maintenance quantization for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7362–7371, 2024. 6

    2024

  7. [15]

    Automatic tuberculosis screening using chest radio- graphs

    Stefan Jaeger, Alexandros Karargyris, Sema Candemir, Les Folio, Jenifer Siegelman, Fiona Callaghan, Zhiyun Xue, Kannappan Palaniappan, Rahul K Singh, Sameer Antani, et al. Automatic tuberculosis screening using chest radio- graphs. IEEE Transactions on Medical Imaging, 33(2):2...

    2013

  8. [16]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 2023. 1, 3

    2023

  9. [17]

    Learning what not to segment: A new perspective on few- shot segmentation

    Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few- shot segmentation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 8057–8067, 2022. 3

    2022

  10. [18]

    Base and meta: A new perspective on few-shot segmentation

    Chunbo Lang, Gong Cheng, Binfei Tu, Chao Li, and Jun- wei Han. Base and meta: A new perspective on few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10669–10686, 2023. 1, 3

    2023

  11. [19]

    Cross-domain few-shot se- mantic segmentation

    Shuo Lei, Xuchao Zhang, Jianfeng He, Fanglan Chen, Bowen Du, and Chang-Tien Lu. Cross-domain few-shot se- mantic segmentation. In European Conference on Computer Vision, pages 73–90. Springer, 2022. 2, 3, 6, 7

    2022

  12. [20]

    Adaptive prototype learning and allocation for few-shot segmentation

    Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8334–8343, 2021. 3

    2021

  13. [21]

    Fss-1000: A 1000-class dataset for few- shot segmentation

    Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few- shot segmentation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 2869–2878, 2020. 6

    2020

  14. [22]

    Inter- mediate prototype mining transformer for few-shot semantic segmentation

    Yuanwei Liu, Nian Liu, Xiwen Yao, and Junwei Han. Inter- mediate prototype mining transformer for few-shot semantic segmentation. Advances in Neural Information Processing Systems, 35:38020–38031, 2022. 1, 3

    2022

  15. [23]

    Simpler is better: Few-shot semantic seg- mentation with classifier weight transformer

    Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic seg- mentation with classifier weight transformer. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 8741–8750, 2021. 3

    2021

  16. [24]

    Segment anything 9 model for medical image analysis: an experimental study

    Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything 9 model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. 1

    2023

  17. [25]

    Hypercorre- lation squeeze for few-shot segmentation

    Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorre- lation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 6941–6952, 2021. 3, 7

    2021

  18. [26]

    Feature weighting and boosting for few-shot segmentation

    Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 622–631, 2019. 6

    2019

  19. [27]

    Cross-domain few-shot segmentation via iterative support-query correspon- dence mining

    Jiahao Nie, Yun Xing, Gongjie Zhang, Pei Yan, Aoran Xiao, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Cross-domain few-shot segmentation via iterative support-query correspon- dence mining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pa...

    2024

  20. [28]

    A threshold selection method from gray-level histograms

    Nobuyuki Otsu et al. A threshold selection method from gray-level histograms. Automatica, 11(285-296):23–27,

  21. [29]

    Hierarchical dense cor- relation distillation for few-shot segmentation

    Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense cor- relation distillation for few-shot segmentation. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23641–23651, 2023. 1, 3

    2023

  22. [30]

    Segment anything meets point tracking

    Frano Raji ˇc, Lei Ke, Yu-Wing Tai, Chi-Keung Tang, Mar- tin Danelljan, and Fisher Yu. Segment anything meets point tracking. arXiv preprint arXiv:2307.01197, 2023. 2

    2023 arXiv

  23. [31]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 3, 6, 8

    2024 arXiv

  24. [32]

    One-shot learning for semantic segmentation

    Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410, 2017. 6

    2017 arXiv

  25. [33]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4

    2010 arXiv

  26. [34]

    Score-based generative modeling through stochastic differential equa- tions

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. arXiv preprint arXiv:2011.13456, 2020. 3

    2011 arXiv

  27. [35]

    Domain-rectifying adapter for cross-domain few-shot segmentation

    Jiapeng Su, Qi Fan, Wenjie Pei, Guangming Lu, and Fanglin Chen. Domain-rectifying adapter for cross-domain few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24036– 24045, 2024. 2, 3, 7

    2024

  28. [36]

    Vrp-sam: Sam with visual reference prompt

    Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. Vrp-sam: Sam with visual reference prompt. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 23565–23574, 2024. 2, 3, 7

    2024

  29. [37]

    Prior guided feature enrich- ment network for few-shot segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence , 44(2):1050– 1065, 2020

    Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrich- ment network for few-shot segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence , 44(2):1050– 1065, 2020. 3, 7

    2020

  30. [38]

    The ham10000 dataset, a large collection of multi-source der- matoscopic images of common pigmented skin lesions

    Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source der- matoscopic images of common pigmented skin lesions. Sci- entific data, 5(1):1–9, 2018. 6

    2018

  31. [39]

    Panet: Few-shot image semantic seg- mentation with prototype alignment

    Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic seg- mentation with prototype alignment. In proceedings of the IEEE/CVF International Conference on Computer Vision , pages 9197–9206, 2019. 7

    2019

  32. [40]

    Remember the differ- ence: Cross-domain few-shot semantic segmentation via meta-memory transfer

    Wenjian Wang, Lijuan Duan, Yuxi Wang, Qing En, Jun- song Fan, and Zhaoxiang Zhang. Remember the differ- ence: Cross-domain few-shot semantic segmentation via meta-memory transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7065...

    2022

  33. [41]

    Rethinking the correlation in few-shot segmentation: A buoys view

    Yuan Wang, Rui Sun, and Tianzhu Zhang. Rethinking the correlation in few-shot segmentation: A buoys view. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7183–7192, 2023. 1, 3

    2023

  34. [42]

    Accelerating diffu- sion sampling with optimized time steps

    Shuchen Xue, Zhaoqiang Liu, Fei Chen, Shifeng Zhang, Tianyang Hu, Enze Xie, and Zhenguo Li. Accelerating diffu- sion sampling with optimized time steps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8292–8301, 2024. 6

    2024

  35. [43]

    Prototype mixture models for few-shot semantic segmentation

    Boyu Yang, Chang Liu, Bohao Li, Jianbin Jiao, and Qix- iang Ye. Prototype mixture models for few-shot semantic segmentation. In Computer Vision–ECCV 2020: 16th Euro- pean Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part VIII 16, pages 763–778. Springer, 2020. 7

    2020

  36. [44]

    Track anything: Segment anything meets videos

    Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 2

    2023 arXiv

  37. [45]

    Pyramid graph networks with connection attentions for region-based one-shot semantic segmentation

    Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid graph networks with connection attentions for region-based one-shot semantic segmentation. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision , pages 9587–9595,

  38. [46]

    Canet: Class-agnostic segmentation networks with it- erative refinement and attentive few-shot learning

    Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with it- erative refinement and attentive few-shot learning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5217–5226, 2019. 7

    2019

  39. [47]

    A comprehensive survey on segment anything model for vision and beyond

    Chunhui Zhang, Li Liu, Yawen Cui, Guanjie Huang, Weilin Lin, Yiqian Yang, and Yuehong Hu. A comprehensive survey on segment anything model for vision and beyond. arXiv preprint arXiv:2305.08196, 2023. 1

    2023 arXiv

  40. [48]

    Few-shot segmentation via cycle-consistent trans- former

    Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent trans- former. Advances in Neural Information Processing Systems, 34:21984–21996, 2021. 3

    2021

  41. [49]

    mixup: Beyond empirical risk minimization

    Hongyi Zhang. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 6

    2017 arXiv

  42. [50]

    Diffmorpher: Unleashing the capability of dif- fusion models for image morphing

    Kaiwen Zhang, Yifan Zhou, Xudong Xu, Bo Dai, and Xin- gang Pan. Diffmorpher: Unleashing the capability of dif- fusion models for image morphing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7912–7921, 2024. 2, 3, 4, 8 10

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.