Pith. sign in

REVIEW 5 major objections 5 minor 111 references

SPARED: Reasoning-Based AI-Generated Image Detection via Adversarially Edited Data

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

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

Pith's one-line read A detector trained by an adversarial editor that forges paired fakes from real photos improves monotonically and beats models 26 times larger.

desk verdict SPARED is a solid, honestly written adversarial-training paper for explainable deepfake detection whose main weaknesses are unshipped artifacts and a shortcut-risk that is reduced but not mathematically closed. read the letter →

arxiv 2608.12876 v1 pith:D2WCMZGV submitted 2026-08-13 cs.CV cs.AI

Yicheng Bao , Xiahui Guo , Xuhong Wang , Xin Tan This is my paper · ORCID
classification cs.CVcs.AI
keywords AI-generatedimagedetectionadversarialreinforcementlearningdiffusioneditingdeepfakemultimodalLLMreasoningexplainableforensicsdatasetbiasGRPO
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 introduces SPARED, an adversarial reinforcement-learning loop in which one model edits real photographs into fake counterparts and a second, reasoning-based model is trained to catch them. The central claim is that this paired editing setup removes the usual dataset-bias shortcut—real and fake images differing by provenance rather than by generation trace—while the alternation keeps forcing the detector to face newly generated hard cases. The authors report monotonic improvement across three external benchmarks, with a 9-billion-parameter detector surpassing much larger closed and open models on judge-style detection and semantic-anomaly reasoning, and explanations improving even though only verdict accuracy is rewarded.

What carries the argument

The load-bearing object is a three-channel adversarial loop. The attacker is a diffusion image editor trained with an RL objective; its reward is $r_A = r_{\mathrm{det}}$ if an instruction-following gate (PaCo) scores the edit at or above 0.7, and 0 otherwise, so a rollout must both obey the edit instruction and fool the current frozen defender to earn credit. The defender is a multimodal large language model fine-tuned with GRPO under the reward $\mathbf{1}[\hat{y}=y]$, which never scores the explanation text. The data channel pairs every edited fake with its own real source image, filtered for deduplication and perceptual-hash leakage against evaluation benchmarks. The definition of the PaCo gate is doing the anti-collapse work: without it, an attacker rewarded only for fooling drifts toward tiny edits, and the resulting pool teaches the defender to call realistic images fake.

What would settle it

Train the same loop on a paired pool in which every fake is also systematically recompressed or brightness-shifted by a fixed amount while still passing the fidelity gate, then test the resulting detector on unedited real images with the same recompression or brightness shift. If accuracy on those shifted real images rises as much as it does on genuine edits, the loop is exploiting the side effect rather than the editing trace; if it does not, the paired-source assumption is doing the claimed work.

Watch

Extended reading notes

Core claim

SPARED claims that a detector can be made both generalizing and explainable without any reward for explanation quality, provided the training loop is shortcut-proof on all three channels: the defender's reward credits only the final real/fake verdict, the attacker's reward is gated on instruction-faithful edits so it cannot win by not editing, and every fake is paired with its own real source so provenance cannot substitute for the editing trace. Under this loop, each attacker round forges a harder pool aimed at the current defender's blind spots, and each defender round is trained on that pool with GRPO. The paper reports that accuracy improves monotonically from SFT through three adversarial iterations on DeepfakeJudge-Detect, AnomReason-Deepfake, and Holmes-Set; the final 9B model surpasses every non-reasoning MLLM evaluated and every reasoning model up to 30B on the first, trailing only a reasoning model 26 times larger, and reaches 92.8 mean accuracy zero-shot on ten unseen generator families. Explanations rise in semantic quality as a side effect of verdict-only training, and ablations show that removing the gate, unpairing the reals, or replacing the evolving attacker with static or fresh data all substantially reduce the gain.

Load-bearing premise

The whole design rests on the gate correctly identifying when an edit is faithful, and on the paired-source construction leaving the editing trace as the only systematic difference between real and fake training images; if the editing model introduces a consistent side effect such as a global brightness shift or compression signature, the defender can learn that shortcut and the reported gains would transfer that shortcut rather than true detection ability.

Editorial extensions

If this is right

  • A detector trained this way should keep improving as generators change, because each round's attacker regenerates the pool against the current decision boundary rather than against a fixed corpus.
  • The verdict-only reward suggests that explanations can be treated as byproducts of accuracy; systems that want trustworthy rationales may not need explanation-specific supervision.
  • The ablations imply that static training, longer training, and even fresh data cannot substitute for an evolving attacker, so future detectors should budget compute for online adversarial regeneration.
  • The zero-shot transfer from locally edited photos to fully synthetic images from unseen generator families indicates that edit-faithful hard negatives teach features that are shared across synthesis pipelines.

Reading between the lines

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

  • Beyond the paper's benchmarks, the same paired-edit loop could be applied to other forensics tasks—document forgery, audio deepfakes, or tampered video—wherever a faithful edit can be generated alongside its pristine source.
  • The reported Janus regression suggests the binary fooling reward has no per-family difficulty control; a graded or calibrated reward could allow the loop to allocate difficulty without sacrificing one generator family while improving the mean.
  • If the efficiency result is taken seriously, it implies parameter count is not the binding constraint for detection; the binding constraint is the difficulty distribution of the training data, which the attacker controls.
  • A direct testable consequence is that replaying the loop with a weaker attacker should lower the ceiling: comparing two checkpoints of the same attacker with different edit diversity would isolate how much detector gain is attributable to pool diversity versus pool difficulty.
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 / 5 minor

Summary. The paper introduces SPARED, an alternating attacker-defender reinforcement-learning loop for explainable AI-generated-image (AIGI) detection. A diffusion image editor (Qwen-Image-Edit LoRA, trained with DiffusionNFT) edits real photographs into paired fakes and is rewarded only when the edit passes a PaCo instruction-fidelity gate and fools a frozen defender; a Qwen3.5-9B reasoning MLLM is trained with verdict-only GRPO on the regenerated pools. The authors claim three 'shortcut-proof' locks: a verdict-only defender reward, a gated attacker reward, and strict pairing of each fake with its own real source. They report monotonic improvements across iterative rounds on three external benchmarks (DeepfakeJudge-Detect, AnomReason-Deepfake, and Holmes-Set), including zero-shot transfer on the latter two, and matched-budget ablations that separate static-pool, fresh-source, ungated, and unpaired training conditions.

Significance. If the shortcut-proof design is established, this is a valuable contribution: it addresses the dataset-bias shortcut at the data-construction level via paired editing rather than post-hoc alignment, and it provides a clean demonstration that explanation quality can rise as a side effect of accuracy-only RL. The ablation design is a real strength: the static-pool, fresh-source, ungated, and unpaired controls directly test several alternative explanations for the gains, and the authors honestly report a per-generator regression (Janus). However, the central claim that the loop escalates genuine detection capability rests on the assertion that PaCo-gated edits differ from their sources only by the instructed editing trace, and this premise is not yet verified. The missing variance estimates and the deferred technical appendix also prevent the experimental claims from being fully audited.

major comments (5)
  1. [Data Construction and Training Schedule; Eq. (3)] The claim that paired editing leaves 'the only systematic difference' between real and fake training images is asserted, not established. PaCo scores instruction-following; it does not constrain output-side low-level statistics. A diffusion editing pipeline can impose a reproducible side channel across all x_edit outputs—for example, a VAE decoding signature, fixed output resolution, resampling/JPEG re-encoding, brightness/gamma shifts, or sharpening. If such a channel exists, a defender trained on paired data can separate the two classes with that cue alone, and the reported benchmark gains would reflect transfer of the shortcut rather than explanation-grounded detection. The unpaired ablation in Table 3 does not settle this: replacing each fake's paired real with a disjoint-source real changes both the pairing and the source distribution, adding a provenance confound. Please add a direct test: train a simple probe classifier on low-level statistics of paired source/edit images (e.g., frequency spectra, color histograms, compression residuals, VAE reconstruction error) and report whether it can separate them; also include a null-edit/identity control to confirm that the PaCo gate rejects unedited images. This is load-bearing for the paper's 'shortcut-proof by design' claim.
  2. [Tables 1-3 and Ablation Study] No error bars, seeds, or repeated runs are reported for any training stage. Both the GRPO defender and the DiffusionNFT attacker are stochastic RL procedures, and several key comparisons in Table 3 are small (e.g., 73.7 vs 73.6 on DFJ-Detect; 0.4998 vs 0.4990 CSemAP-Full). The monotonicity and matched-budget claims need variance estimates; at minimum, report three seeds per condition with the range or standard deviation, and state whether the reported models are the best or the average checkpoint.
  3. [Experimental Setup and Conclusion] The manuscript repeatedly defers to a 'technical appendix' and 'supplementary material' for corpus sizes, hyperparameters, filtering and screening thresholds, the per-generator Holmes-Set table, the GenShield comparison, and the UniGenDet Holmes-Set number (99.2). None of this material is included in the submission. Without it, the matched-budget ablations and the training configuration cannot be audited, and the paper is not reproducible. The appendix should be part of the submission, or the missing numbers should be moved into the main text.
  4. [Reasoning Defender, Eq. (1)] The statement that 'the explanation is never rewarded' is inaccurate as written. The LoRA-SFT initialization trains directly on real/fake reasoning pairs from DeepfakeJudge, so the explanation text is supervised before the GRPO phase; only the later RL phase restricts reward to verdict correctness. This matters for the templated-rationale argument, because SFT on a fixed explanation corpus could itself imprint template rationales. Please rephrase to 'never rewarded in the RL phase' and discuss what the SFT phase contributes to the CSemAP gains observed in Table 2.
  5. [Eq. (3) and Data Construction and Training Schedule] The PaCo gate threshold (0.7) and the number of adversarial rounds are free hyperparameters, and no sensitivity analysis is reported. The monotonicity result is specific to these choices: a higher threshold could gate out all adversarial examples, while a lower threshold could admit degenerate edits. Please report a small sweep over the gate threshold, and show what happens when the loop is stopped earlier or extended beyond Iter3.
minor comments (5)
  1. [Figure 3 and per-generator table] The caption states that the final round's AP stays above 92 on every Holmes-Set family, but the per-generator values are only in the missing appendix; key values should be moved into the main text or an accessible supplementary file.
  2. [Method and Introduction] The model name 'Qwen-Image-Edit-2511' appears in the Method, while the introduction and the DeepfakeJudge-Detect description refer to 'Qwen-Edit'. Please align the naming and verify the intended model identifier.
  3. [Abstract and Conclusion] The phrase 'the explanation is never rewarded' should be qualified as 'never rewarded in the RL phase' to avoid a direct contradiction with the supervised SFT initialization described in the Method section.
  4. [Experimental Setup] The statement that all three benchmarks show monotonic improvement is slightly overstated for DeepfakeJudge-Detect, since the SFT initialization uses that suite's designated training corpus and the benchmark is therefore not fully zero-shot; the Experimental Setup acknowledges this, but the abstract and conclusion should as well.
  5. [Throughout] There are several PDF-extraction spacing errors in the text (e.g., 'templatedrationales', 'trainedwithaccuracy-basedGRPO'). Please ensure the final version is typeset correctly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the training loop, external evaluations, and ablations keep the central claims independent of their inputs.

full rationale

SPARED's derivation chain is self-contained rather than circular. The defender is trained with a verdict-only reward (Eq. 1) and the attacker with a PaCo-gated fooling reward (Eqs. 2-3); neither target quantity is defined in terms of the reported benchmark metrics. The paper's central claims—monotonic improvement across three external benchmarks, rising explanation quality despite no explanation reward, and generalization to unseen generators—are all evaluated on externally curated test sets (DeepfakeJudge-Detect, AnomReason-Deepfake, Holmes-Set) that are not used to fit the reported model. The ablations in Table 3 directly test the load-bearing design choices (static pool, frozen attacker, missing gate, unpaired reals) and show that removing each component eliminates the gain, which is evidence that the observed improvement is not merely a re-description of the training objective. The PaCo gate is an external reward model, and the paper's assumption that paired editing leaves only the editing trace as the systematic difference is an empirical premise subject to side-channel risk, but that is a correctness or generalization concern, not circularity: the paper does not define 'detection capability' in terms of its own training signal, nor does it fit a parameter and then relabel that fit as a prediction. Self-citations to Tan et al. 2025 and related prior work are used as benchmark references and background, not as load-bearing justification for the paper's own conclusions. Therefore no specific circular step can be exhibited, and the honest finding is no significant circularity.

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

The central claims rest on the empirical reliability of the PaCo instruction-fidelity gate, the paired-source data construction closing provenance shortcuts, the verdict-only reward producing emergent explanation quality, and the perceptual-hash screening preventing benchmark leakage. These are domain assumptions with no formal or independent verification in the paper.

free parameters (2)
  • PaCo gate threshold = 0.7
    Hand-chosen threshold in Eq. (3) for the instruction-fidelity gate; determines which attacker rollouts receive reward and therefore shapes the entire training pool.
  • Number of adversarial rounds = 5 total; Iter3 reported as final
    Schedule choice: the paper reports Iter1-3 as the defender stages and claims monotonic improvement, but does not analyze sensitivity to the number of rounds or the stop at round 3.
assumptions (4)
  • domain assumption The PaCo reward model reliably distinguishes instruction-faithful edits from unfaithful ones at the 0.7 threshold.
    The whole attacker reward (Eq. 3) depends on r_PaCo >= 0.7; if the gate is gameable or miscalibrated, the loop can reward degenerate edits.
  • domain assumption Paired-source editing leaves the editing trace as the only systematic class-conditional difference in the training pool.
    The claim that dataset-bias shortcuts are closed by construction relies on this; any consistent non-editing side effect of the edit model would give the defender a shortcut.
  • domain assumption Verdict-only GRPO reward causes explanation quality to improve as a side effect.
    The paper predicts this from the reward design and observes it, but treats it as a consequence with no external guarantee.
  • domain assumption The perceptual-hash screening prevents training-source leakage into all three evaluation benchmarks.
    Screening thresholds are in the technical appendix and not verified; leakage would inflate zero-shot numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SPARED: Reasoning-Based AI-Generated Image Detection via Adversarially Edited Data." pith.science (2026). https://pith.science/paper/D2WCMZGV

@misc{pith2026260812876,
  author       = {Pith},
  title        = {Pith review of: SPARED: Reasoning-Based AI-Generated Image Detection via Adversarially Edited Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D2WCMZGV}},
  note         = {Machine review of arXiv:2608.12876}
}
read the original abstract

Detecting AI-generated images is only half the task: a deployed detector must also justify its verdict, yet existing detectors inherit three failure modes from their training data: real and fake images collected from different sources invite provenance shortcuts, supervised explanation corpora teach templated rationales, and a static forgery corpus leaves the decision boundary standing still while generators keep moving. We introduce \methodname{}, an adversarial reinforcement learning framework that pits two heterogeneous models against each other. A diffusion image editor learns to edit real photographs into fake counterparts of those same photographs that fool the current detector, while a reasoning MLLM learns to expose them with a verdict grounded in free-form reasoning. Both rewards are shortcut-proof by design: the attacker is credited only when its edit is faithfully executed, and the defender only when its verdict is correct. As the two models alternate, each round's attacker regenerates a harder training pool aimed at the current detector's blind spots, so the detector must generalize rather than memorize any fixed artifact distribution. Although the explanation is never rewarded, its quality rises round over round as a side effect of accuracy-only training. A detector trained within this loop improves monotonically across rounds on each of three external benchmarks.

Figures

Figures reproduced from arXiv: 2608.12876 by the authors.

Figure 1
Figure 1. (a) A detector trained once on a static corpus is a sta [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of SPARED: one adversarial round, read left to right. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a, b) Per-generator accuracy and AP on Holmes-Set across training stages, with axes ordered by final-round accuracy: [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative cases. (a) A mechanical violation: the base model judges the image real and explicitly reports no [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

111 extracted references · 53 canonical work pages

  1. [1]

    2026 , doi =

    Yuxuan Chou and Tao Yu and Wen Huang and Yuheng Zhang and Tao Dai and Shu-Tao Xia , booktitle =. 2026 , doi =

    2026

  2. [2]

    2022 , url =

    Liang Chen and Yong Zhang and Yibing Song and Lingqiao Liu and Jue Wang , booktitle =. 2022 , url =

    2022

  3. [3]

    2024 , eprint =

    Yuzhen Lin and Wentang Song and Bin Li and Yuezun Li and Jiangqun Ni and Han Chen and Qiushi Li , booktitle =. 2024 , eprint =

    2024

  4. [4]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages =

    Muzhi Dai and Shixuan Liu and Zhiyuan Zhao and Junyu Gao and Hao Sun and Xuelong Li , year =. Proceedings of the 33rd ACM International Conference on Multimedia , pages =

  5. [5]

    Xiaoyu Wen and Zhida He and Han Qi and Ziyu Wan and Zhongtian Ma and Ying Wen and Tianhang Zheng and Xingcheng Xu and others , year =

  6. [6]

    Mickel Liu and Liwei Jiang and Yancheng Liang and Simon Shaolei Du and Yejin Choi and Tim Althoff and Natasha Jaques , year =

  7. [7]

    Ziyao Huang and Weiwei Wu and Kui Wu and Jianping Wang and Wei-Bin Lee , year =

  8. [8]

    Yuxiang Wei and Zhiqing Sun and Emily McMilin and Jonas Gehring and David Zhang and Gabriel Synnaeve and Daniel Fried and Lingming Zhang and others , year =

Show all 111 references
  1. [9]

    Chengsong Huang and Wenhao Yu and Xiaoyang Wang and Hongming Zhang and Zongxia Li and Ruosen Li and Jiaxin Huang and Haitao Mi and others , year =

  2. [10]

    2024 , howpublished =

    2024

  3. [11]

    International Conference on Machine Learning , year =

    Patrick Esser and Sumith Kulal and Andreas Blattmann and Rahim Entezari and Jonas M. International Conference on Machine Learning , year =

  4. [12]

    2021 , editor =

    Alec Radford and Jong Wook Kim and Chris Hallacy and Aditya Ramesh and Gabriel Goh and Sandhini Agarwal and Girish Sastry and Amanda Askell and others , booktitle =. 2021 , editor =

    2021

  5. [13]

    Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Xiao Bi and Haowei Zhang and Mingchuan Zhang and others , journal =

  6. [14]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle =

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle =. 2022 , url =

    2022

  7. [15]

    2508.02324 , archivePrefix =

    Chenfei Wu and Jiahao Li and Jingren Zhou and Junyang Lin and Kaiyuan Gao and Kun Yan and Sheng-ming Yin and Shuai Bai and others , year =. 2508.02324 , archivePrefix =

    arXiv

  8. [16]

    Kaiwen Zheng and Huayu Chen and Haotian Ye and Haoxiang Wang and Qinsheng Zhang and Kai Jiang and Hang Su and Stefano Ermon and others , journal =

  9. [17]

    2512.04784 , archivePrefix =

    Bowen Ping and Chengyou Jia and Minnan Luo and Changliang Xia and Xin Shen and Zhuohang Dang and Hangwei Qian , year =. 2512.04784 , archivePrefix =

  10. [18]

    2505.20275 , archivePrefix =

    Yang Ye and Xianyi He and Zongjian Li and Bin Lin and Shenghai Yuan and Zhiyuan Yan and Bohan Hou and Li Yuan , year =. 2505.20275 , archivePrefix =

    arXiv

  11. [19]

    2510.19808 , archivePrefix =

    Yusu Qian and Eli Bocek-Rivele and Liangchen Song and Jialing Tong and Yinfei Yang and Jiasen Lu and Wenze Hu and Zhe Gan , year =. 2510.19808 , archivePrefix =

  12. [20]

    2023 , url =

    Kai Zhang and Lingbo Mo and Wenhu Chen and Huan Sun and Yu Su , booktitle =. 2023 , url =

    2023

  13. [21]

    2511.21631 , archivePrefix =

    Shuai Bai and Yuxuan Cai and Ruizhe Chen and Keqin Chen and Xionghui Chen and Zesen Cheng and Lianghao Deng and Wei Ding and others , year =. 2511.21631 , archivePrefix =

    arXiv

  14. [22]

    2020 , journal =

    Joel Frank and Thorsten Eisenhofer and Lea Sch. 2020 , journal =

    2020

  15. [23]

    Efros , year =

    Sheng-Yu Wang and Oliver Wang and Richard Zhang and Andrew Owens and Alexei A. Efros , year =. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    2020

  16. [24]

    Nan Zhong and Yiran Xu and Sheng Li and Zhenxing Qian and Xinpeng Zhang , year =

  17. [25]

    Chuangchuang Tan and Huan Liu and Yao Zhao and Shikui Wei and Guanghua Gu and Ping Liu and Yunchao Wei , year =

  18. [26]

    2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Utkarsh Ojha and Yuheng Li and Yong Jae Lee , year =. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    2023

  19. [27]

    2025 , url =

    Dimitrios Karageorgiou and Symeon Papadopoulos and Ioannis Kompatsiaris and Efstratios Gavves , booktitle =. 2025 , url =

    2025

  20. [28]

    2025 , url =

    Jeongsoo Park and Andrew Owens , booktitle =. 2025 , url =

    2025

  21. [29]

    2407.20836 , archivePrefix =

    Yunfeng Diao and Naixin Zhai and Changtao Miao and Zitong Yu and Xingxing Wei and Xun Yang and Meng Wang , year =. 2407.20836 , archivePrefix =

  22. [30]

    2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Fabrizio Guillaro and Giada Zingarini and Ben Usman and Avneesh Sud and Davide Cozzolino and Luisa Verdoliva , year =. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    2025

  23. [31]

    2025 , eprint =

    Ruoxin Chen and Junwei Xi and Zhiyuan Yan and Ke-Yue Zhang and Shuang Wu and Jingyi Xie and Xu Chen and Lei Xu and others , booktitle =. 2025 , eprint =

    2025

  24. [32]

    2025 , url =

    Anirudh Sundara Rajan and Utkarsh Ojha and Jedidiah Schloesser and Yong Jae Lee , booktitle =. 2025 , url =

    2025

  25. [33]

    doi:10.48550/arXiv.2605.27348 , url =

    Jihyeon Kim and Sohee Kim and Soosan Lee and Souhwan Jung and James Matthew Rehg and Hyesong Choi , year =. doi:10.48550/arXiv.2605.27348 , url =. 2605.27348 , archivePrefix =

    doi arXiv

  26. [34]

    Shilin Yan and Ouxiang Li and Jiayin Cai and Yanbin Hao and Xiaolong Jiang and Yao Hu and Weidi Xie , year =

  27. [35]

    Zhiyuan Yan and Jiangming Wang and Peng Jin and Ke-Yue Zhang and Chengchun Liu and Shen Chen and Taiping Yao and Shouhong Ding and others , year =

  28. [36]

    Jiazhen Yan and Ziqiang Li and Fan Wang and Boyu Wang and Ziwen He and Zhangjie Fu , year =

  29. [37]

    Jiazhen Yan and Fan Wang and Weiwei Jiang and Ziqiang Li and Zhangjie Fu , year =

  30. [38]

    doi:10.48550/arXiv.2509.19841 , url =

    Tai-Ming Huang and Wei-Tung Lin and Kai-Lung Hua and Wen-Huang Cheng and Junichi Yamagishi and Jun-Cheng Chen , year =. doi:10.48550/arXiv.2509.19841 , url =. 2509.19841 , archivePrefix =

    doi

  31. [39]

    2025 , publisher =

    Tianxiao Li and Zhenglin Huang and Haiquan Wen and Yiwei He and Shuchang Lyu and Baoyuan Wu and Guangliang Cheng , booktitle =. 2025 , publisher =. doi:10.1145/3746027.3754798 , url =

    2025

  32. [40]

    2026 , url =

    Yikun Ji and Yan Hong and Qi Fan and Jun Lan and Huijia Zhu and Weiqiang Wang and Liqing Zhang and Jianfu Zhang , booktitle =. 2026 , url =

    2026

  33. [41]

    2026 , doi =

    OuCheng Huang and Manxi Lin and Jiexiang Tan and Xiaoxiong Du and Yang Qiu and Junjun Zheng and Xiangheng Kong and Yuning Jiang and others , booktitle =. 2026 , doi =

    2026

  34. [42]

    Kartik Kuckreja and Parul Gupta and Muhammad Haris Khan and Abhinav Dhall , year =

  35. [43]

    2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Zhenglin Huang and Jinwei Hu and Xiangtai Li and Yiwei He and Xingyu Zhao and Bei Peng and Baoyuan Wu and Xiaowei Huang and others , year =. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    2025

  36. [44]

    2025 IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

    Ziyin Zhou and Yunpeng Luo and Yuanchen Wu and Ke Sun and Jiayi Ji and Ke Yan and Shouhong Ding and Xiaoshuai Sun and others , year =. 2025 IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

    2025

  37. [45]

    Chuangchuang Tan and Xiang Ming and Jinglu Wang and Renshuai Tao and Bin Li and Yunchao Wei and Yao Zhao and Yan Lu , year =

  38. [46]

    Siwei Wen and Junyan Ye and Peilin Feng and Hengrui Kang and Zichen Wen and Yize Chen and Jiang Wu and Wenjun Wu and others , year =

  39. [47]

    Yangjun Wu and Keyu Yan and Yu Liu and Jingren Zhou and Fei Huang and Rong Zhang and Zhou Zhao and Fei Wu , year =

  40. [48]

    Yanran Zhang and Wenzhao Zheng and Yifei Li and Bingyao Yu and Yu Zheng and Lei Chen and Jiwen Lu and Jie Zhou , year =

  41. [49]

    Zhipei Xu and Xuanyu Zhang and Youmin Xu and Qing Huang and Shen Chen and Taiping Yao and Shouhong Ding and Jian Zhang , year =

  42. [50]

    doi:10.48550/arXiv.2507.06261 , url =

    Comanici, Gheorghe and Bieber, Eric and Schaekermann, Mike and others , year =. doi:10.48550/arXiv.2507.06261 , url =. 2507.06261 , archivePrefix =

    doi arXiv

  43. [51]

    2024 , month = aug, url =

    2024

  44. [52]

    2025 , month = apr, url =

    2025

  45. [53]

    2025 , month = aug, url =

    Introducing. 2025 , month = aug, url =

    2025

  46. [54]

    2024 , month = jul, url =

    2024

  47. [55]

    2024 , eprint =

    Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling , author =. 2024 , eprint =

    2024

  48. [56]

    doi:10.48550/arXiv.2504.10479 , url =

    Zhu, Jinguo and Wang, Weiyun and Chen, Zhe and others , year =. doi:10.48550/arXiv.2504.10479 , url =. 2504.10479 , archivePrefix =

    doi arXiv

  49. [57]

    doi:10.48550/arXiv.2508.18265 , url =

    Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and others , year =. doi:10.48550/arXiv.2508.18265 , url =. 2508.18265 , archivePrefix =

    doi arXiv

  50. [58]

    doi:10.48550/arXiv.2409.12191 , url =

    Wang, Peng and Bai, Shuai and Tan, Sinan and others , year =. doi:10.48550/arXiv.2409.12191 , url =. 2409.12191 , archivePrefix =

    doi arXiv

  51. [59]

    doi:10.48550/arXiv.2502.13923 , url =

    Bai, Shuai and Chen, Keqin and Liu, Xuejing and others , year =. doi:10.48550/arXiv.2502.13923 , url =. 2502.13923 , archivePrefix =

    doi arXiv

  52. [60]

    2024 , eprint =

    Long Context Transfer from Language to Vision , author =. 2024 , eprint =

    2024

  53. [61]

    doi:10.48550/arXiv.2408.03326 , url =

    Li, Bo and others , year =. doi:10.48550/arXiv.2408.03326 , url =. 2408.03326 , archivePrefix =

    doi arXiv

  54. [62]

    doi:10.48550/arXiv.2404.14219 , url =

    Abdin, Marah and Jacobs, Sam Ade and Awan, Ammar Ahmad and others , year =. doi:10.48550/arXiv.2404.14219 , url =. 2404.14219 , archivePrefix =

    doi arXiv

  55. [63]

    doi:10.48550/arXiv.2408.01800 , url =

    Yao, Yuan and others , year =. doi:10.48550/arXiv.2408.01800 , url =. 2408.01800 , archivePrefix =

    doi arXiv

  56. [64]

    doi:10.48550/arXiv.2408.04840 , url =

    Ye, Jiabo and others , year =. doi:10.48550/arXiv.2408.04840 , url =. 2408.04840 , archivePrefix =

    doi arXiv

  57. [65]

    doi:10.48550/arXiv.2412.08905 , url =

    Abdin, Marah and others , year =. doi:10.48550/arXiv.2412.08905 , url =. 2412.08905 , archivePrefix =

    doi arXiv

  58. [66]

    doi:10.48550/arXiv.2507.14632 , url =

    Wen, Haiquan and Li, Tianxiao and Huang, Zhenglin and He, Yiwei and Cheng, Guangliang , year =. doi:10.48550/arXiv.2507.14632 , url =. 2507.14632 , archivePrefix =

    doi arXiv

  59. [67]

    doi:10.48550/arXiv.2310.17419 , url =

    Chang, You-Ming and Yeh, Chen and Chiu, Wei-Chen and Yu, Ning , year =. doi:10.48550/arXiv.2310.17419 , url =. 2310.17419 , archivePrefix =

    doi arXiv

  60. [68]

    doi:10.48550/arXiv.2403.17465 , url =

    Luo, Yunpeng and Du, Junlong and Yan, Ke and Ding, Shouhong , year =. doi:10.48550/arXiv.2403.17465 , url =. 2403.17465 , archivePrefix =

    doi arXiv

  61. [69]

    2024 , eprint =

    Leveraging Representations from Intermediate Encoder-Blocks for Synthetic Image Detection , author =. 2024 , eprint =

    2024

  62. [70]

    doi:10.48550/arXiv.2410.13848 , url =

    Wu, Chengyue and others , year =. doi:10.48550/arXiv.2410.13848 , url =. 2410.13848 , archivePrefix =

    doi arXiv

  63. [71]

    doi:10.48550/arXiv.2501.17811 , url =

    Chen, Xiaokang and others , year =. doi:10.48550/arXiv.2501.17811 , url =. 2501.17811 , archivePrefix =

    doi arXiv

  64. [72]

    doi:10.48550/arXiv.2408.12528 , url =

    Xie, Jinheng and others , year =. doi:10.48550/arXiv.2408.12528 , url =. 2408.12528 , archivePrefix =

    doi arXiv

  65. [73]

    2024 , eprint =

    Autoregressive Model Beats Diffusion: Llama for Scalable Image Generation , author =. 2024 , eprint =

    2024

  66. [74]

    doi:10.48550/arXiv.2412.04431 , url =

    Han, Jian and others , year =. doi:10.48550/arXiv.2412.04431 , url =. 2412.04431 , archivePrefix =

    doi arXiv

  67. [75]

    2024 , eprint =

    Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction , author =. 2024 , eprint =

    2024

  68. [76]

    doi:10.48550/arXiv.2310.00426 , url =

    Chen, Junsong and others , year =. doi:10.48550/arXiv.2310.00426 , url =. 2310.00426 , archivePrefix =

    doi arXiv

  69. [77]

    Scott McCloskey and Michael Albright , year =

  70. [78]

    Xiuli Bi and Bo Liu and Fan Yang , journal =

  71. [79]

    Riccardo Corvi and Davide Cozzolino and Giada Zingarini , booktitle =

  72. [80]

    Zhendong Wang and Jianmin Bao and Wengang Zhou , booktitle =

  73. [81]

    Mingjian Zhu and Hanting Chen and Mouxiao Huang , year =

  74. [82]

    Mingjian Zhu and Hanting Chen and Qiangyu Yan , booktitle =

  75. [83]

    Bin Cao and Jianhao Yuan and Yexin Liu , year =

  76. [84]

    Jiaxuan Chen and Jieteng Yao and Li Niu , year =

  77. [85]

    Yize Chen and Zhiyuan Yan and Guangliang Cheng , year =

  78. [86]

    Sungik Choi and Hankook Lee and Jaehoon Lee , year =

  79. [87]

    Zhiyuan He and Pin-Yu Chen and Tsung-Yi Ho , year =

  80. [88]

    Zhengchao Huang and Bin Xia and Zicheng Lin , year =

  81. [89]

    Huan Liu and Zichang Tan and Chuangchuang Tan , booktitle =

  82. [90]

    Zihan Liu and Hanyi Wang and Yaoyu Kang , year =

  83. [91]

    Jonas Ricker and Denis Lukovnikov and Asja Fischer , booktitle =

  84. [92]

    Chuangchuang Tan and Renshuai Tao and Huan Liu , booktitle =

  85. [93]

    Zhipei Xu and Xuanyu Zhang and Runyi Li , booktitle =

  86. [94]

    Junyan Ye and Baichuan Zhou and Zilong Huang , year =

  87. [95]

    Fanrui Zhang and Jiawei Liu and Jiaying Zhu , year =

  88. [96]

    Siyuan Cheng and Lingjuan Lyu and Zhenting Wang , booktitle =

  89. [97]

    Davide Cozzolino and Giovanni Poggi and Matthias Nießner , booktitle =

  90. [98]

    Wan Jiang and Jing Yan and Xiaojing Chen , year =

  91. [99]

    Hengrui Kang and Siwei Wen and Zichen Wen , booktitle =

  92. [100]

    Yixuan Li and Xuelin Liu and Xiaoyang Wang , journal =

  93. [101]

    Yiheng Li and Zichang Tan and Guoqing Xu , year =

  94. [102]

    Shuqiao Liang and Jian Liu and Renzhang Chen , year =

  95. [103]

    Ruiqi Liu and Yi Han and Zhengbo Zhang , year =

  96. [104]

    Ziheng Qin and Yuheng Ji and Renshuai Tao , year =

  97. [105]

    Zheng Yang and Ruoxin Chen and Zhiyuan Yan , year =

  98. [106]

    Peipeng Yu and Jianwei Fei and Hui Gao , booktitle =

  99. [107]

    Yanran Zhang and Bingyao Yu and Yu Zheng , booktitle =

  100. [108]

    Lei Tan and Shuwei Li and Mohan Kankanhalli , booktitle =

  101. [109]

    Xinchang Wang and Yunhao Chen and Yuechen Zhang , year =

  102. [110]

    Wenbin Wang and Yuge Huang and Jianqing Xu , year =

  103. [111]

    Haiwei Wu and Jiantao Zhou and Shile Zhang , journal =

Pith tools

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