Pith. sign in

REVIEW 4 major objections 7 minor 2 cited by

Interpretable and Reliable Detection of AI-Generated Images via Grounded Reasoning in MLLMs

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning a vision-language model on point-and-explain annotations yields a detector that reports 98.1% accuracy and localizes the flaws it cites.

desk verdict A genuinely new bbox-plus-text dataset for explainable AIGC detection, with a plausible but unpolished training pipeline; a misprinted reward formula, an inconsistent IoU headline, and loose annotation QC need fixing before this is benchmark-ready. read the letter →

arxiv 2506.07045 v1 pith:FNV23KQE submitted 2025-06-08 cs.CV cs.AIcs.CL

Yikun Ji , Hong Yan , Jun Lan , Huijia Zhu , Weiqiang Wang , Qi Fan , Liqing Zhang , Jianfu Zhang This is my paper
classification cs.CVcs.AIcs.CL
keywords AI-generatedimagedetectionexplainablemultimodallargelanguagemodelsvisualgroundingGRPOgroundedreasoningFakeXplaineddatasethallucinationreduction
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 tries to establish that fine-tuning a multimodal large language model on human-annotated bounding boxes and captions of synthesis artifacts turns it into a detector that says not just whether an image is AI-generated but where and why it looks fake. On the FakeXplained dataset, the fine-tuned model reaches 98.1% classification accuracy and a localization IoU of 37.8%, outperforming classification-only and segmentation-based baselines and transferring to unseen generators. The reason to care is practical: a binary verdict is hard to trust or audit, whereas a grounded explanation gives users, journalists, and platforms a way to verify the model's judgment against the image itself.

What carries the argument

The load-bearing construction is FakeXplained plus a two-stage training protocol. FakeXplained supplies the ground-truth link between a visible region and a human reason; the protocol turns that link into a reward signal using a structured output with <think>, <tag>, and <verdict> markers. Supervised fine-tuning first teaches the model the output format, then three Group Relative Policy Optimization (GRPO) stages progressively reweight rewards for format compliance, correct label, and relaxed Intersection over Union, so the final stage pushes the model to localize artifacts precisely without sacrificing detection.

What would settle it

Re-annotate a random sample of FakeXplained images with independent expert annotators under strict agreement (for example, over 50% IoU and full tag consensus), then measure the fine-tuned model's IoU against that stricter ground truth; a large drop from 37.8% would show that the reported localization is calibrated to permissive annotation noise rather than to objective artifacts. A complementary check would run the human preference study against two independent human annotation sets to see whether human-human agreement exceeds human-model agreement.

Watch

Extended reading notes

Core claim

The paper's central claim is that fine-grained human-aligned grounding supervision is what makes MLLM-based detection reliable rather than hallucinated. Each of the 8,772 fake images in FakeXplained carries, on average, 5.42 pairs of bounding boxes and descriptive captions pointing at visual flaws, plus image-level perceptual tags; training on these turns Qwen-2.5-VL-32B into a detector that emits a structured answer with reasoned regions, tags, and a verdict. The authors report that this grounded model beats all tested baselines on accuracy and localization, reaches near-human preference in a pairwise study against human annotations, and keeps most of its performance on FaceForensics++ and ChatGPT-4o-generated images.

Load-bearing premise

Human annotations of 'fake regions' are accurate enough to serve as ground truth: only 5% of images were quality-checked, with tolerances of 20% box overlap and 33.3% tag accuracy, so noisy or inconsistent references would inflate both training supervision and reported IoU.

Editorial extensions

If this is right

  • If the paper is right, a detector for AI-generated images can be both accurate and explainable: 98.1% accuracy and human-comparable captions are obtained in one end-to-end model.
  • Fine-grained grounding appears to be a source of generalization: the model beats baselines on out-of-distribution data including FaceForensics++ and ChatGPT-4o images.
  • Reasoning supervision itself contributes to accuracy: training on binary labels alone reaches 93.5%, whereas the full grounded pipeline reaches 98.1%.
  • The learned cues are semantic rather than pixel-level, since accuracy holds under JPEG compression, random cropping, and downsampling.

Reading between the lines

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

  • If annotation noise is as high as the quality-control thresholds suggest (20% box IoU, 33.3% tag accuracy, 5% of images checked), the reported 37.8% IoU could overstate true localization; re-annotation under strict expert agreement would calibrate it.
  • The same two-stage grounded-reasoning recipe could be transplanted to other human-verifiable visual tasks, such as medical imaging and document forgery, wherever point-and-explain annotations are available.
  • The absolute IoU of 37.8% remains modest even if it beats baselines, so a natural next test is whether raising the QC threshold changes the gap between model and human annotations.
  • An untested extension is ablating boxes versus captions per generator family; the paper reports overall averages, but the relative value of grounding may differ between diffusion models and GANs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper presents FakeXplained, a new human-annotated dataset of 8,772 AI-generated images with bounding boxes and captions describing synthesis artifacts, and trains Qwen-2.5-VL-32B via SFT followed by three GRPO stages with rewards for classification accuracy, a relaxed IoU for localization, and output format validity. The resulting model is reported to reach 98.1% classification accuracy and an IoU of 37.8% on the FakeXplained test split, outperform segmentation-based baselines on localization, generalize to two out-of-distribution datasets, and produce region-grounded explanations that approach human annotation quality in a preference study. The authors claim this is the first dataset to provide localized, textual explanations for AI-generated image detection and that their progressive GRPO strategy is essential to the improvement.

Significance. If the results hold, the paper makes a useful contribution to explainable AI-generated image detection: it provides a sizable grounded dataset, demonstrates that a large MLLM can be tuned to output both a classification verdict and localized natural-language reasons, and includes extensive comparisons, ablations, robustness tests, and a human preference study. The dataset alone is a potentially valuable resource for the community, and the multi-stage SFT+GRPO recipe for grounding is a plausible template for other perception tasks. The paper also ships a broad ablation matrix across data components, training strategies, model sizes, and base models. However, the central claims currently rest on several load-bearing technical details that are either misstated or underspecified: the printed grounding reward formula is mathematically inverted, the headline IoU numbers disagree across tables, and the localization evaluation protocol is not fully defined. These issues prevent the reader from confirming that the reported localization gains are real and attributable to the proposed reward.

major comments (4)
  1. [§4.3, Eq. (3)] The relaxed IoU reward is written as RG(o) = IoU × η = min(1, η |R(o) ∪ Ry| / |R(o) ∩ Ry|). As printed, the union/intersection ratio is inverted: for any pair of non-empty boxes, |R(o) ∪ Ry| ≥ |R(o) ∩ Ry|, so the ratio is at least 1 and, with η ≥ 1 as suggested by Table 1 ('IoU×1.1', '1.0×', '1.5×', '2.0×'), the min term equals 1 regardless of overlap; if the intersection is empty the ratio is infinite and the reward is still 1. Thus the formula as written supplies no localization signal (and would actually reward disjoint boxes). Please correct the equation to the intended form (presumably min(1, η · |R(o) ∩ Ry| / |R(o) ∪ Ry|)) and state whether the implementation used the corrected formula. If the implementation used the printed formula, the observed IoU improvements could not have come from RG, which would invalidate the attribution of the GRPO gains to grounding.
  2. [§5.2, Tables 2 and 4] The headline localization result is inconsistent across the manuscript. Section 5.2 states that the best model achieves an IoU score of 37.8%, and Table 4's 'Best' column reports IoU 0.378, but Table 2's overall row for the same method reports IoU 0.352 (and the per-generator rows range from 0.330 to 0.368). Since Table 2 is the main comparison against baselines, the reader cannot tell whether the claimed superiority over ObjectFormer (0.299) and SegFormer (0.289) is based on 0.352 or 0.378. Please reconcile these numbers, identify which evaluation protocol produces each value, and state which is the definitive result. In addition, Table 4 shows GRPOγ reaching IoU 0.401, higher than the best model's 0.378, yet the progressive pipeline is declared superior; without error bars or significance tests, this trade-off is not established.
  3. [§3.2, Quality control] The reliability of the FakeXplained annotations is a load-bearing assumption because the same annotations are used both as SFT/GRPO supervision and as the IoU evaluation target. The quality-control protocol validates only 5% of images, requires only 20% IoU for spatial agreement with a reference box, and only 33.3% accuracy for image-level tags, and no inter-annotator agreement is reported. With an average of 5.42 boxes per image and 23 annotators, the absence of pairwise IoU or tag-agreement statistics leaves open the possibility that the supervision is noisy or idiosyncratic, in which case the 37.8% IoU measures agreement with one annotation protocol rather than accurate localization of genuine artifacts. Please report inter-annotator agreement on a random subset (e.g., mean pairwise IoU, tag Cohen's kappa or Fleiss' kappa), describe how the 5% validation set was selected, and discuss how the adopted thresholds affect the interpretation of the localization numbers.
  4. [§5.2, IoU evaluation protocol] The paper reports a single IoU number per model but does not specify how predicted bounding boxes are matched to the ground-truth boxes, despite images having on average 5.42 annotated regions. Different matching rules (e.g., greedy matching, Hungarian matching, or per-ground-truth maximum IoU) can produce substantially different scores, and a permissive rule could inflate the reported IoU. Please specify the exact matching algorithm, whether multiple predictions are allowed to match the same ground-truth box, how false-positive and duplicate boxes are penalized, and how real images (with no ground-truth boxes) are excluded from the metric. Without this information the localization comparison against segmentation baselines is not reproducible.
minor comments (7)
  1. [§2] The phrase 'remains challenging DefakeByReals' is incomplete and the reference is not properly cited; please fix the sentence and add the intended citation.
  2. [Table 3] The table title says 'unseen image categories' but the rows are the out-of-distribution datasets OpenAI 4o and FaceForensics++; please align the title with the content.
  3. [Table 2] In the LlamaGen row, ObjectFormer reports IoU 0.429, which is far above its own range (0.275–0.312) and above the best method's overall value; please verify that this entry is not a typo.
  4. [§4.3] The notation R(o) is used both for the region extraction function and for the predicted region set, and Ry is written inconsistently with the earlier Rr in §3.2; please use distinct symbols and define the domain of each.
  5. [§5.2, Human preference] The preference study reports 1,525 valid votes after removing neutral responses but does not state the number of evaluators, the number of images, or the inter-rater reliability; please provide these details so the near-parity conclusion can be assessed.
  6. [All experimental tables] No standard deviations or confidence intervals are reported despite four-fold cross-validation; please include error bars at least for the headline accuracy and IoU numbers in Tables 2 and 4.
  7. [§4.3] The reward design does not include any term that measures caption quality or region-caption alignment, yet the abstract and introduction claim the pipeline balances 'coherent textual explanation'; please either add such a reward or soften the claim to reflect that caption quality is only supervised via SFT.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are supported by held-out evaluation, external OOD benchmarks, and standard supervised fine-tuning rather than by self-referential derivation.

full rationale

The paper is an empirical training-and-evaluation study, not a derivation from first principles. The central claims—98.1% classification accuracy and 37.8% IoU—are measured on held-out splits via four-fold cross-validation, with the IoU computed against human-annotated boxes that are not used to fit the evaluated model on that fold. The model is additionally tested on two external datasets (FaceForensics++ and ChatGPT-4o images), so the reported generalization is not forced by construction. The reward functions for GRPO do use human-annotated labels, including a relaxed IoU term, but optimizing a model against a training signal and then measuring the same quantity on held-out data is standard supervised learning, not circular reasoning. The human-preference study compares model outputs against human annotations from the dataset, but the evaluators are independent and the comparison is an empirical quality check rather than a logical reduction. There are no load-bearing self-citations: the cited methods (GRPO, Qwen-2.5-VL, R1-V, etc.) are external prior work with no author overlap that would make the argument depend on the present authors' own unverified claims. The paper does contain internal inconsistencies and potential correctness concerns—the stated relaxed-IoU formula appears to invert the IoU ratio, and the reported overall IoU differs between Table 2 (0.352) and the text/Table 4 (0.378)—but these are technical errors or reporting mismatches, not circularity. Likewise, the reliance on human annotations in FakeXplained, with QC performed on only 5% of images at a 20% IoU threshold, is a data-quality and validity concern, not a self-referential derivation; the evaluation remains external to the model's own outputs. Overall, no prediction or claimed result reduces by construction to its own input, and no self-citation chain is load-bearing. The appropriate circularity score is therefore 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claims depend on the quality and representativeness of the human annotations, on the curated dataset being a fair proxy for AIGC detection, and on the unstated corrected version of the grounding reward in Eq. (3). There are no invented physical entities. The paper's own extra beyond standard components is the annotated dataset and hand-set reward weights.

free parameters (6)
  • η (relaxed IoU constant) = 1.1 (per text; stages scale reward by 1.0x-2.0x)
    Hand-chosen constant in the grounding reward RG (Eq. 3). The printed formula min(1, η|R∪Ry|/|R∩Ry|) is always 1 for any non-empty overlap when η>1, so the actual implemented form must differ from what is written.
  • GRPO reward weights (rbase, ωG, ωC, ωF) = Table 1: Base 0/-0.5/-1; Label +1/−1, +2/−2, 0.5/−1; Format +2/−1, +1/−1.5, 0.5/−1 across stages α/β/γ
    Hand-set stage weights for the three GRPO stages; no automated tuning or validation-based selection is reported.
  • Number of SFT epochs = 3
    Chosen by monitoring convergence of the SFT loss (Appendix A.2).
  • Learning rates = 1e-4 (SFT), 1e-5 (RLHF)
    Stated in Appendix A.2 as initial learning rates; standard choices.
  • Number of chat templates = 9 (3 for real, 6 for generated)
    Chosen to reduce overfitting; Appendix E.3 shows sensitivity on OoD accuracy.
  • Number of GRPO stages and epochs = 3 stages (α, β, γ), 1 epoch each
    Stage count and schedule chosen by design; ablation compares against fixed-weight variants.
assumptions (4)
  • domain assumption Human annotations of fake regions are valid ground truth for synthesis artifacts.
    Used as SFT supervision and as the IoU evaluation target. Validated on only 5% of images with an IoU threshold of 20% and tag accuracy threshold of 33.3% (Section 3.2).
  • domain assumption The curated FakeXplained set of 8,772 images from 28 generators is representative enough to train a generalizable detector.
    Images were top-rated by volunteers and selected for annotatable flaws; OoD tests on ChatGPT-4o and FaceForensics++ only partially cover high-fidelity fakes without obvious artifacts.
  • ad hoc to paper The GRPO grounding reward, as implemented, provides a meaningful learning signal for box overlap.
    Eq. (3) as printed is mathematically inverted and caps at 1 for any non-empty overlap, so the paper relies on an unstated corrected reward function.
  • domain assumption SFT followed by three GRPO stages with the stated reward weights converges to a stable policy.
    Empirical claim based on Figure 3; no statistical guarantee is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpretable and Reliable Detection of AI-Generated Images via Grounded Reasoning in MLLMs." pith.science (2026). https://pith.science/paper/FNV23KQE

@misc{pith2026250607045,
  author       = {Pith},
  title        = {Pith review of: Interpretable and Reliable Detection of AI-Generated Images via Grounded Reasoning in MLLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FNV23KQE}},
  note         = {Machine review of arXiv:2506.07045}
}
read the original abstract

The rapid advancement of image generation technologies intensifies the demand for interpretable and robust detection methods. Although existing approaches often attain high accuracy, they typically operate as black boxes without providing human-understandable justifications. Multi-modal Large Language Models (MLLMs), while not originally intended for forgery detection, exhibit strong analytical and reasoning capabilities. When properly fine-tuned, they can effectively identify AI-generated images and offer meaningful explanations. However, existing MLLMs still struggle with hallucination and often fail to align their visual interpretations with actual image content and human reasoning. To bridge this gap, we construct a dataset of AI-generated images annotated with bounding boxes and descriptive captions that highlight synthesis artifacts, establishing a foundation for human-aligned visual-textual grounded reasoning. We then finetune MLLMs through a multi-stage optimization strategy that progressively balances the objectives of accurate detection, visual localization, and coherent textual explanation. The resulting model achieves superior performance in both detecting AI-generated images and localizing visual flaws, significantly outperforming baseline methods.

Figures

Figures reproduced from arXiv: 2506.07045 by the authors.

Figure 1
Figure 1. An overview of our method. The rise of Multi-modal Large Language Models (MLLMs) has enabled cross-modal inference, allowing models to generate human-readable reasoning about AI-generated images [8, 39, 50]. Recent efforts [27, 67] have advanced interpretable textual explanations using MLLMs. However, without visual grounding, it remains uncertain whether these rationales accurately reflect image content or derive f… view at source ↗
Figure 3
Figure 3. Accuracy, IoU metric (upper), loss and reward [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. A sample user query and the corresponding model output (visualized). [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: An example showing two different chat templates branched from one annotation entry. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: More annotation examples from FakeXplained and model response visualized from the [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. XPlainVerse: A Million-Scale Benchmark for Explainable Deepfake Detection

    cs.CV 2026-07 conditional novelty 6.5 of 10

    A million-scale deepfake benchmark with Edit-Check filtering, dual expert/lay explanations, and EntityScore/EvidenceScore shows fine-tuned detectors collapse under generator shift while surface fluency remains.

  2. AgentFoX: LLM Agent-Guided Fusion with eXplainability for AI-Generated Image Detection

    cs.CV 2026-03 conditional novelty 6.0 of 10

    An LLM agent guided by Expert and Clustering Profiles fuses heterogeneous AIGI detectors, resolves conflicts, and outputs explainable forensic reports that beat single experts and standard ensembles on high-conflict a...

Reference graph

Works this paper leans on

68 extracted references · 22 canonical work pages · cited by 2 Pith papers

  1. [1]

    Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C. Cosman. Detecting generated images by real images only, 2023

    work page 2023

  2. [2]

    Large scale gan training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. InInternational Conference on Learning Representations, 2018

    work page 2018

  3. [3]

    Maskgit: Masked generative image transformer

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11315–11325, 2022

    2022

  4. [4]

    Antifakeprompt: Prompt-tuned vision-language models are fake image detectors.arXiv preprint arXiv:2310.17419, 2023

    You-Ming Chang, Chen Yeh, Wei-Chen Chiu, and Ning Yu. Antifakeprompt: Prompt-tuned vision-language models are fake image detectors.arXiv preprint arXiv:2310.17419, 2023

    arXiv 2023

  5. [5]

    Pixart- σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation

    Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. InEuropean Conference on Computer Vision, pages 74–91. Springer, 2024

    2024

  6. [6]

    Pixart- δ: Fast and controllable image generation with latent consistency models.arXiv preprint arXiv:2401.05252, 2024

    Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart- δ: Fast and controllable image generation with latent consistency models.arXiv preprint arXiv:2401.05252, 2024

    arXiv 2024

  7. [7]

    Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis.arXiv preprint arXiv:2310.00426, 2023

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis.arXiv preprint arXiv:2310.00426, 2023

    arXiv

  8. [8]

    R1-v: Reinforcing super generaliza- tion ability in vision-language models with less than $3.https://github.com/Deep-Agent/ R1-V, 2025

    Liang Chen, Lei Li, Haozhe Zhao, Yifan Song, and Vinci. R1-v: Reinforcing super generaliza- tion ability in vision-language models with less than $3.https://github.com/Deep-Agent/ R1-V, 2025. Accessed: 2025-02-02

    2025

Show all 68 references
  1. [9]

    Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling, 2025

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, Lixin Gu, Xuehui Wang, Qingyun Li, Yimin Ren, Zixuan Chen, Jiapeng Luo, Jiahao Wang, Tan Jiang, Bo Wang, Conghui He, Botian Shi, Xingcheng Zhang, Han Lv, Yi...

    2025

  2. [10]

    On the detection of synthetic images generated by diffusion models

    Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. InIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, 2023

    2023

  3. [11]

    Li, and Li Fei-Fei

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database.2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009

    2009

  4. [12]

    Witherden

    Tarik Dzanic, Karan Shah, and Freddie D. Witherden. Fourier spectrum discrepancies in deep network generated images. InProceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20, Red Hook, NY , USA, 2020. Curran Associates Inc

    2020

  5. [13]

    Scaling rectified flow transformers for high-resolution image synthesis, 2024

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transf...

    2024

  6. [14]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873–12883, 2021. 17

    2021

  7. [15]

    Leveraging frequency analysis for deep fake image recognition

    Joel Frank, Thorsten Eisenhofer, Lea Schönherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. InProceedings of the 37th International Conference on Machine Learning, ICML’20. JMLR.org, 2020

    2020

  8. [16]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Y . Bengio. Generative adversarial networks.Advances in Neural Information Processing Systems, 3, 06 2014

    2014

  9. [17]

    Fake or jpeg? revealing common biases in generated image detection datasets.ArXiv, abs/2403.17608, 2024

    Patrick Grommelt, Louis Weiss, Franz-Josef Pfreundt, and Janis Keuper. Fake or jpeg? revealing common biases in generated image detection datasets.ArXiv, abs/2403.17608, 2024

    2024 arXiv

  10. [18]

    Vector quantized diffusion model for text-to-image synthesis

    Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696–10706, 2022

    2022

  11. [19]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

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

    2025 arXiv

  12. [20]

    Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024

    2024 arXiv

  13. [21]

    Jonathan Ho, Ajay Jain, and P. Abbeel. Denoising diffusion probabilistic models.ArXiv, abs/2006.11239, 2020

    2006 arXiv

  14. [22]

    Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749, 2025

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749, 2025

    2025 arXiv

  15. [23]

    Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    2024 arXiv

  16. [24]

    T2i-r1: Reinforcing image generation with collaborative semantic-level and token-level cot.arXiv preprint arXiv:2505.00703, 2025

    Dongzhi Jiang, Ziyu Guo, Renrui Zhang, Zhuofan Zong, Hao Li, Le Zhuo, Shilin Yan, Pheng- Ann Heng, and Hongsheng Li. T2i-r1: Reinforcing image generation with collaborative semantic-level and token-level cot.arXiv preprint arXiv:2505.00703, 2025

    2025 arXiv

  17. [25]

    A style-based generator architecture for generative adversarial networks.2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396–4405, 2018

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks.2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396–4405, 2018

    2019

  18. [26]

    Flux.https://github.com/black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/black-forest-labs/flux, 2024

    2024

  19. [27]

    Fakebench: Uncover the achilles’ heels of fake images with large multimodal models.ArXiv, abs/2404.13306, 2024

    Yixuan Li, Xuelin Liu, Xiaoyang Wang, Shiqi Wang, and Weisi Lin. Fakebench: Uncover the achilles’ heels of fake images with large multimodal models.ArXiv, abs/2404.13306, 2024

    2024 arXiv

  20. [28]

    Zhengzhe Liu, Xiaojuan Qi, and Philip H.S. Torr. Global texture enhancement for fake face detection in the wild. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8057–8066, 2020

    2020

  21. [29]

    Peter Lorenz, Ricard Durall, and Janis Keuper. Detecting images generated by deep diffusion models using their local intrinsic dimensionality.2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 448–459, 2023

    2023

  22. [30]

    Midjourney, 2023

    Midjourney. Midjourney, 2023

    2023

  23. [31]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021

    2021 arXiv

  24. [32]

    Towards universal fake image detectors that generalize across generative models.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480–24489, 2023

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480–24489, 2023. 18

    2023

  25. [33]

    Dall·e 3 system card, 2023

    OpenAI. Dall·e 3 system card, 2023

    2023

  26. [34]

    Introducing 4o image generation, Mar 2025

    OpenAI. Introducing 4o image generation, Mar 2025

    2025

  27. [35]

    Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730...

    2022

  28. [36]

    Medvlm-r1: Incentivizing medical reasoning capability of vision-language models (vlms) via reinforcement learning.arXiv preprint arXiv:2502.19634, 2025

    Jiazhen Pan, Che Liu, Junde Wu, Fenglin Liu, Jiayuan Zhu, Hongwei Bran Li, Chen Chen, Cheng Ouyang, and Daniel Rueckert. Medvlm-r1: Incentivizing medical reasoning capability of vision-language models (vlms) via reinforcement learning.arXiv preprint arXiv:2502.19634, 2025

    2025 arXiv

  29. [37]

    Community forensics: Using thousands of generators to train fake image detectors, 2024

    Jeongsoo Park and Andrew Owens. Community forensics: Using thousands of generators to train fake image detectors, 2024

    2024

  30. [38]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  31. [39]

    Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl.arXiv preprint arXiv:2503.07536, 2025

    Yingzhe Peng, Gongrui Zhang, Miaosen Zhang, Zhiyuan You, Jie Liu, Qipeng Zhu, Kai Yang, Xingzhong Xu, Xin Geng, and Xu Yang. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl.arXiv preprint arXiv:2503.07536, 2025

    2025 arXiv

  32. [40]

    Qwen2.5-vl, January 2025

    Qwen Team. Qwen2.5-vl, January 2025

    2025

  33. [41]

    Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.06125, 1(2):3, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.06125, 1(2):3, 2022

    2022 arXiv

  34. [42]

    Rapidata openai 4o preference, Mar 2025

    Rapidata. Rapidata openai 4o preference, Mar 2025

    2025

  35. [43]

    Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9130–9140, 2024

    2024

  36. [45]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, June 2022

    2022

  37. [46]

    FaceForensics++: Learning to detect manipulated facial images

    Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics++: Learning to detect manipulated facial images. InInternational Conference on Computer Vision (ICCV), 2019

    2019

  38. [47]

    Photorealistic text-to-image diffusion models with deep language understanding.Advances in Neural Information Processing Systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in Neural Information ...

    2022

  39. [48]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient- based localization. In2017 IEEE International Conference on Computer Vision (ICCV), pages 618–626, 2017

    2017

  40. [49]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Jun-Mei Song, Mingchuan Zhang, Y . K. Li, Yu Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.ArXiv, abs/2402.03300, 2024. 19

    2024 arXiv

  41. [50]

    Vlm-r1: A stable and generalizable r1-style large vision-language model.arXiv preprint arXiv:2504.07615, 2025

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. Vlm-r1: A stable and generalizable r1-style large vision-language model.arXiv preprint arXiv:2504.07615, 2025

    2025 arXiv

  42. [51]

    Deep inside convolutional networks: Visualising image classification models and saliency maps.CoRR, abs/1312.6034, 2013

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps.CoRR, abs/1312.6034, 2013

    2013 arXiv

  43. [52]

    Denoising diffusion implicit models.ArXiv, abs/2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.ArXiv, abs/2010.02502, 2020

    2010 arXiv

  44. [53]

    Autoregressive model beats diffusion: Llama for scalable image generation.arXiv preprint arXiv:2406.06525, 2024

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation.arXiv preprint arXiv:2406.06525, 2024

    2024 arXiv

  45. [54]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning - Volume 70, ICML’17, page 3319–3328. JMLR.org, 2017

    2017

  46. [55]

    Mingxing Tan and Quoc V . Le. Efficientnet: Rethinking model scaling for convolutional neural networks.ArXiv, abs/1905.11946, 2019

    1905 arXiv

  47. [56]

    Galip: Generative adversarial clips for text-to-image synthesis

    Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. Galip: Generative adversarial clips for text-to-image synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14214–14223, 2023

    2023

  48. [57]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in neural information processing systems, 37:84839–84865, 2024

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in neural information processing systems, 37:84839–84865, 2024

    2024

  49. [58]

    Neural discrete representation learning.Advances in neural information processing systems, 30, 2017

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning.Advances in neural information processing systems, 30, 2017

    2017

  50. [59]

    Vl- rethinker: Incentivizing self-reflection of vision-language models with reinforcement learning

    Haozhe Wang, Chao Qu, Zuming Huang, Wei Chu, Fangzhen Lin, and Wenhu Chen. Vl- rethinker: Incentivizing self-reflection of vision-language models with reinforcement learning. arXiv preprint arXiv:2504.08837, 2025

    2025 arXiv

  51. [60]

    Objectformer for image manipulation detection and localization

    Junke Wang, Zuxuan Wu, Jingjing Chen, Xintong Han, Abhinav Shrivastava, Ser-Nam Lim, and Yu-Gang Jiang. Objectformer for image manipulation detection and localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    2022

  52. [61]

    Qwen2-vl: Enhancing vision- language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision- language model’s p...

    2024 arXiv

  53. [62]

    Cnn- generated images are surprisingly easy to spot

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn- generated images are surprisingly easy to spot... for now. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695–8704, 2020

    2020

  54. [63]

    Dire for diffusion-generated image detection

    Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445–22455, 2023

    2023

  55. [64]

    Fast-slow thinking for large vision-language model reasoning.arXiv preprint arXiv:2504.18458, 2025

    Wenyi Xiao, Leilei Gan, Weilong Dai, Wanggui He, Ziwei Huang, Haoyuan Li, Fangxun Shu, Zhelun Yu, Peng Zhang, Hao Jiang, et al. Fast-slow thinking for large vision-language model reasoning.arXiv preprint arXiv:2504.18458, 2025

    2025

  56. [65]

    Segformer: Simple and efficient design for semantic segmentation with transformers

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. InNeural Information Processing Systems (NeurIPS), 2021. 20

    2021

  57. [66]

    R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization.arXiv preprint arXiv:2503.10615, 2025

    Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, et al. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization.arXiv preprint arXiv:2503.10615, 2025

    2025 arXiv

  58. [67]

    A-bench: Are lmms masters at evaluating ai-generated images?arXiv preprint arXiv:2406.03070, 2024

    Zicheng Zhang, Haoning Wu, Chunyi Li, Yingjie Zhou, Wei Sun, Xiongkuo Min, Zijian Chen, Xiaohong Liu, Weisi Lin, and Guangtao Zhai. A-bench: Are lmms masters at evaluating ai-generated images?arXiv preprint arXiv:2406.03070, 2024

    2024 arXiv

  59. [68]

    Swift: a scalable lightweight infrastructure for fine-tuning

    Yuze Zhao, Jintao Huang, Jinghan Hu, Xingjun Wang, Yunlin Mao, Daoze Zhang, Zeyinzi Jiang, Zhikai Wu, Baole Ai, Ang Wang, et al. Swift: a scalable lightweight infrastructure for fine-tuning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 2973...

    2025

  60. [69]

    aha moment

    Hengguang Zhou, Xirui Li, Ruochen Wang, Minhao Cheng, Tianyi Zhou, and Cho-Jui Hsieh. R1-zero’s" aha moment" in visual reasoning on a 2b non-sft model.arXiv preprint arXiv:2503.05132, 2025. 21

    2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.