Pith. sign in

REVIEW 5 major objections 6 minor 68 references

Synthesizing Near-Boundary OOD Samples for Out-of-Distribution Detection

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that near-boundary OOD samples can be synthesized by iterative diffusion inpainting guided by an energy-score gradient, and that fine-tuning CLIP on these samples yields average AUROC 97.01 and FPR95 14.27 on ImageNet…

desk verdict A practical, well-ablated synthetic-data pipeline for CLIP OOD detection whose benchmark gains may be partly driven by semantic overlap between the generated scenes and the test OOD sets. read the letter →

arxiv 2507.10225 v3 pith:F6XWELXK submitted 2025-07-14 cs.CV

Jinglun Li , Kaixun Jiang , Zhaoyu Chen , Bo Lin , Yao Tang , Weifeng Ge , Wenqiang Zhang This is my paper
classification cs.CV
keywords out-of-distributiondetectionCLIPfine-tuningsyntheticdatagenerationnear-boundaryOODsamplesdiffusioninpaintingmultimodallargelanguagemodelsenergyscorenegativelabels
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

SynOOD claims that the hardest failures in CLIP-based out-of-distribution detection, samples that look almost like in-distribution data, can be manufactured rather than collected. The paper builds a pipeline that asks a multimodal language model to list the background context of an ImageNet image, then repeatedly inpaints with a diffusion model while pushing the result toward the InD/OOD decision boundary using an energy-score gradient, producing synthetic near-boundary OOD images. Fine-tuning the CLIP image encoder's projection layer and a subset of negative text-label features on these images gives an average AUROC of 97.01 and an FPR95 of 14.27 on four OOD test sets, beating prior CLIP-based detectors with under 1% added parameters and under 2 ms added inference time. If this holds, difficult OOD data does not have to be awaited in the wild; it can be generated on demand.

What carries the argument

The mechanism that carries the argument is a closed loop between generation and detection. An MLLM supplies contextual prompts; a latent diffusion model performs iterative inpainting with strength below 1, so the scene's background is preserved while the subject is replaced. The key identity is the gradient update on the initial noise, $\epsilon := \epsilon - r \tilde{\nabla}_\epsilon L_O$, where $L_O$ is the energy score $L_O = m_\text{out} - \tau \log \sum_{i=1}^C e^{g_i(x_\text{syn})/\tau}$ computed on a frozen ResNet50 classifier, and a skip-gradient approximation makes the backpropagation tractable by dropping the Jacobian through the diffusion U-Net. This loop pushes generated images toward the detector's boundary rather than merely sampling random OOD-like images. On the fine-tuning side, the CLIP image encoder stays frozen while a projection layer is trained with a CLIP contrastive loss against negative labels, and then roughly half of the negative-label text features are tuned directly, which the paper argues prevents overfitting and keeps far-OOD sensitivity.

What would settle it

Take the generated synthetic set and measure its OOD score under the final fine-tuned CLIP detector. If the generated images' score distribution is not systematically closer to the InD/OOD threshold than ordinary ImageNet images, the claim that the energy-gradient loop manufactures near-boundary samples is false. A second check: replace the energy score with another OOD score, such as maximum softmax probability or nearest-neighbor distance; if the improvement disappears, boundary-adjacent generation rather than the specific energy loss is the wrong explanation.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that iterative inpainting guided by an OOD-score gradient samples from the InD/OOD boundary, and that fine-tuning CLIP on those samples sharpens the boundary. Concretely, starting from an in-distribution image, the method lets an MLLM name contextual elements (for example, 'bamboo' for a panda image), then repeatedly denoises an inpainting diffusion model with a learnable noise vector updated by the gradient of an energy-based OOD loss from a ResNet50 classifier (Equations 7–10). The resulting synthetic images are visually similar to the originals but semantically outside the label space. Paired with the original images and a set of negative labels, they are used to train a projection layer on the frozen CLIP image encoder and to fine-tune about half of the negative-label text features. Across iNaturalist, SUN, Places, and Texture, the method reports average AUROC 97.01 and FPR95 14.27, improving on NegLabel, CSP, and AdaNeg, with under 2 ms added inference per image.

Load-bearing premise

The method assumes that images pushed to the energy-score boundary of a ResNet50 classifier are representative of the hard out-of-distribution cases the final CLIP detector will encounter, even though the boundary is computed on a different architecture and feature space and the test OOD datasets are unseen.

Editorial extensions

If this is right

  • Synthetic near-boundary data can substitute for real OOD examples in fine-tuning CLIP, so detectors can be improved without waiting for or labeling natural OOD collections.
  • The method adds under 1% parameters and under 2 ms inference time, so the improvement is compatible with deployment on top of existing CLIP-based detectors.
  • The gains transfer across CLIP backbones (ResNet50, ViT-B/32, ViT-B/16), suggesting the fine-tuning recipe is architecture-agnostic.
  • On the OpenOOD benchmark, the approach improves far-OOD detection to 96.21 AUROC and near-OOD to 77.55 AUROC while using only a 50k-image subset, showing the generation pipeline helps where real OOD training data is scarce.
  • Fine-tuning about half of the negative label features, rather than all of them, preserves far-OOD performance while sharpening near-boundary discrimination; exceeding that budget degrades results.

Reading between the lines

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

  • A testable extension is to drive the generation loop with a boundary defined directly in CLIP feature space instead of ResNet50 logits; if the proxy boundary is the active ingredient, CLIP-defined boundaries should match or beat the current results, and if not, the ResNet50 proxy may be supplying useful inductive bias.
  • The same loop could be applied to other closed-set classifiers beyond CLIP, such as medical-image or defect-detection models, wherever near-boundary negatives are scarce; the paper does not test this.
  • If the generated data are close enough to the true boundary, they could also serve as calibration or adversarial-training data, a use the paper leaves implicit.
  • The sensitivity to the number of fine-tuned negative labels (peak at 50k synthetic images, decline at 75k) suggests the method's gains depend on a specific balance; reproducing the exact label-subset selection rule is needed before extrapolating.
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

5 major / 6 minor

Summary. The paper proposes SynOOD, a method that generates synthetic near-boundary out-of-distribution (OOD) images for fine-tuning CLIP-based OOD detectors. An MLLM extracts contextual elements from ImageNet images by prompting for everything except the class subject; an iterative inpainting diffusion process replaces the subject with these context elements, guided by an energy-score loss computed with a ResNet50 classifier. The resulting synthetic OOD images, paired with negative labels, are used to fine-tune a projection layer on the CLIP image encoder and a subset of CLIP text-encoder negative-label features. On ImageNet-1k as in-distribution data, SynOOD reports state-of-the-art average AUROC 97.01 and FPR95 14.27 across iNaturalist, SUN, Places, and Texture, with a smaller-than-1% parameter increase and under 2 ms added inference time. The paper also reports results on the OpenOOD benchmark and across CLIP backbones (ResNet50, ViT-B/32, ViT-B/16), together with ablations of the generation strategy, the amount of synthetic data, and the training scheme.

Significance. If the central claim holds, SynOOD is a practically valuable contribution: it shows that foundation-model-based synthetic data can improve boundary-level OOD discrimination for CLIP, and the paper provides a public code release, ablations of design choices, and results across multiple architectures and evaluation protocols. The improvements over NegLabel on Places and Texture are large (AUROC +5.73 and +5.07 in Table 1), suggesting a real effect. However, the significance is moderated by the concern that the synthetic data may be semantically aligned with the scene/texture test distributions by construction, and by the absence of the directly competing DreamOOD baseline and of error bars. These issues do not invalidate the empirical findings but do affect the strength of the generality claim.

major comments (5)
  1. [Section 3.3, Eq. (3)] The synthetic OOD distribution appears to be semantically constructed to match the test OOD sets. The prompt p_in explicitly asks for 'what else is in the context besides the {class name}', and the iterative inpainting replaces the foreground subject with those contextual elements. For ImageNet, such context elements are typically scene components (sky, water, grass, buildings, textures), and the test sets SUN, Places365, and Texture are drawn from exactly these distributions. The largest gains in Table 1 are on Places (+5.73 AUROC over NegLabel) and Texture (+5.07), consistent with this confound. The paper provides no analysis of the overlap between the MLLM-extracted context terms and the test OOD class semantics, nor an experiment on a held-out OOD set that is not scene/texture-like (e.g., a non-natural-image dataset). Without such evidence, the contribution may be benchmark-specific data augmentation rather than general near-boundary OOD synthesis. I ask the authors to add a control experiment that evaluates on OOD sets with different semantic structure, and to quantify the semantic overlap between p_con and the test OOD datasets.
  2. [Section 3.3, Eqs. (7)-(10)] The near-boundary property is defined by an energy score on a ResNet50 classifier, while the final detector uses CLIP ViT-B/16. The paper does not show that samples pushed toward the ResNet50 energy boundary are also near the CLIP decision boundary, nor that the generated images are actually near any boundary under the final scoring function. The Skip Gradient approximation in Eq. (9) drops the Jacobian terms involving the diffusion denoising process, so the gradient signal is a coarse proxy; the paper claims rapid convergence of L_O but reports no distribution of final energy scores or any verification that the generated samples occupy the intended region. I request an analysis showing the energy scores of synthetic images under both ResNet50 and CLIP (or under the final OOD score used at test time), together with a comparison to real OOD samples, to support the claim that the generated data are genuinely near-boundary for the deployed detector.
  3. [Related Work and Table 1] DreamOOD [62] is cited as a method that learns a text-conditioned latent space to generate diverse OOD samples and is the closest generation-based baseline to SynOOD, yet it is absent from the main comparison in Table 1. Since the paper claims state-of-the-art performance, a direct comparison to DreamOOD on the same four OOD datasets is necessary, or the authors should state clearly why this baseline is excluded. Without this comparison, the superiority claim is not fully established for the generation-based family of methods.
  4. [Section 3.4] The construction of the fine-tuning dataset is under-specified. The paper states that each synthetic image is paired one-to-one with a negative label y^- in D_syn, but it does not specify how a given synthetic image is assigned to a particular negative label from Y^-, nor how the subset Y^*_ of 'associated' negative labels is formed. In addition, the selection of the InD subset D_in^* is based on a JPEG file-size-to-pixel-ratio heuristic, but no ablation compares this heuristic with random selection or with other complexity measures. These choices are load-bearing for the fine-tuning step and for the claim that the negative-label pairing drives the improvement, so they need to be described precisely and validated.
  5. [Table 1] All reported numbers are single-run point estimates without standard deviations or significance tests. Given that the generation process is stochastic (diffusion noise and iterative gradient updates) and the fine-tuning uses only 3-5 epochs, run-to-run variance could be non-negligible; several of the claimed improvements over AdaNeg, for example, are on the order of 1-2 FPR95 points. Please report mean and standard deviation over at least three random seeds for the main experiments, or explain why variance is negligible.
minor comments (6)
  1. [Eq. (2)] The denominator in Eq. (2) appears to be missing a plus sign: it should read sim(x, Y) + sim(x, Y^-) rather than sim(x,Y) sim(x,Y^-).
  2. [Section 3.5, Eqs. (13)-(14)] In Eq. (13) the synthetic image feature is defined as I_syn_i, but Eq. (14) uses I_i in the numerator; please make the notation consistent.
  3. [Table 1 and Section 4.3] There are typos in the headers: 'iNatrualist' should be 'iNaturalist' in Table 1, and 'Anaylsis' should be 'Analysis' in Section 4.3.
  4. [Figure 2] The label 'Label y_i\'' in Figure 2 is ambiguous; it should be clarified whether this is an InD label or a negative label.
  5. [Section 4.1] The paper states that SynOOD adds 'less than 1% additional parameters', but the number of learnable negative-label features in T_neg^* and the projection layer size are not given; please report the exact parameter counts.
  6. [Conclusion] The paper would benefit from a short limitations paragraph, in particular discussing the dependence on the MLLM/diffusion model quality and the computational cost of the offline generation phase.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline results are measured on held-out OOD datasets, and the energy-score guidance is a proxy used only for data synthesis, not the evaluation metric.

full rationale

I walked the derivation chain from Eq. (3) through Eq. (14) and the evaluation tables. The synthetic OOD images are generated by MLLM-context inpainting with an energy-score gradient on a ResNet50 (Eqs. 7-10), but this OOD score is not reused by the final CLIP detector; the detector is evaluated on held-out iNaturalist, SUN, Places, Texture, and OpenOOD sets. No fitted constant is renamed as a prediction: m_out is a synthesis target, and the reported AUROC/FPR95 values are computed from the fine-tuned CLIP score on test data. The learnable negative labels are trained only on synthetic data and tested on unseen OOD images, which is standard supervised fine-tuning rather than circularity. The cited Skip Gradient operation [32] is an external technique adopted for computational efficiency and does not smuggle in the target result. The self-citations [19,20] are contextual related-work citations, not load-bearing premises. The observation that MLLM-extracted scene/texture context may overlap with the SUN/Places/Texture test distributions is a benchmark-validity or distribution-overlap concern, not a circular derivation, because the test images themselves are never used to generate or select the synthetic data. No specific step reduces, by construction or self-citation, to its own inputs; therefore no circularity is established.

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

The method relies on several domain assumptions about the reliability of foundation models and the proxy OOD boundary; no new physical or conceptual entities are introduced. Free parameters are mostly training hyperparameters, but the energy threshold, synthetic dataset size, and number of learnable negative labels directly shape the reported gains.

free parameters (4)
  • Energy threshold m_out = not reported
    Constant in the OOD loss (Eq. 7) taken from Liu et al. [3]; determines how far generated images are pushed toward the proxy boundary.
  • Gradient scale r*rho = 10
    Product of the noise learning rate and the skip-gradient scale (Section 4.1); controls the size of each noise update in the generation loop.
  • Number of synthetic samples m = 50k (main results)
    Size of the paired InD/synthetic dataset used for fine-tuning; Figure 3 shows performance changes with m, so the choice directly affects the headline numbers.
  • Number of fine-tuned negative labels M' = about 5.5k of 11k
    Half of the negative labels are made learnable (Section 3.5); the selection criterion for which labels become learnable is not specified.
assumptions (4)
  • domain assumption The energy score of a ResNet50 classifier is a valid proxy for the InD/OOD boundary.
    Used as the OOD loss to guide noise updates in Section 3.3, Eq. (7); the final detector is CLIP, and the boundary alignment is never validated.
  • domain assumption The MLLM (LLaVA) reliably enumerates contextual elements whose substitution yields meaningful OOD prompts.
    The inpainting prompts are generated entirely by the MLLM (Section 3.3); if these elements are off-target, the synthetic images may not be near-boundary.
  • ad hoc to paper The Skip Gradient approximation from Chen et al. [32] preserves enough gradient signal to guide diffusion noise toward the energy boundary.
    Introduced in Eq. (9) to reduce computational cost; no analysis is provided for how the approximation affects the final image distribution.
  • domain assumption WordNet-derived negative labels (from NegLabel [18]) cover the semantic space of possible OOD categories.
    The fine-tuning contrastive losses align synthetic images to negative labels; if the negative label set misses relevant semantics, boundary alignment could degrade.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Synthesizing Near-Boundary OOD Samples for Out-of-Distribution Detection." pith.science (2026). https://pith.science/paper/F6XWELXK

@misc{pith2026250710225,
  author       = {Pith},
  title        = {Pith review of: Synthesizing Near-Boundary OOD Samples for Out-of-Distribution Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F6XWELXK}},
  note         = {Machine review of arXiv:2507.10225}
}
read the original abstract

Pre-trained vision-language models have exhibited remarkable abilities in detecting out-of-distribution (OOD) samples. However, some challenging OOD samples, which lie close to in-distribution (InD) data in image feature space, can still lead to misclassification. The emergence of foundation models like diffusion models and multimodal large language models (MLLMs) offers a potential solution to this issue. In this work, we propose SynOOD, a novel approach that harnesses foundation models to generate synthetic, challenging OOD data for fine-tuning CLIP models, thereby enhancing boundary-level discrimination between InD and OOD samples. Our method uses an iterative in-painting process guided by contextual prompts from MLLMs to produce nuanced, boundary-aligned OOD samples. These samples are refined through noise adjustments based on gradients from OOD scores like the energy score, effectively sampling from the InD/OOD boundary. With these carefully synthesized images, we fine-tune the CLIP image encoder and negative label features derived from the text encoder to strengthen connections between near-boundary OOD samples and a set of negative labels. Finally, SynOOD achieves state-of-the-art performance on the large-scale ImageNet benchmark, with minimal increases in parameters and runtime. Our approach significantly surpasses existing methods, and the code is available at https://github.com/Jarvisgivemeasuit/SynOOD.

Figures

Figures reproduced from arXiv: 2507.10225 by the authors.

Figure 1
Figure 1. (a) illustrates a simplified example highlighting the lim [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed SynOOD framework. (a) Near-boundary OOD image generation: utilizes an MLLM and a diffusion [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. SynOOD performance with different amounts of syn [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of InD and synthetic OOD data. The labels at the top of the figure represent ImageNet categories as InD. The first [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 54 canonical work pages

  1. [62]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,”Advances in neural information processing systems, vol. 36, 2024. 5

    work page 2024

  2. [1]

    A baseline for detecting misclassified and out-of-distribution examples in neural net- works,

    D. Hendrycks and K. Gimpel, “A baseline for detecting misclassified and out-of-distribution examples in neural net- works,”arXiv preprint arXiv:1610.02136, 2016. 1, 2, 5, 6

    arXiv 2016

  3. [2]

    Enhancing the reliability of out-of-distribution image detection in neural networks,

    S. Liang, Y . Li, and R. Srikant, “Enhancing the reliability of out-of-distribution image detection in neural networks,” arXiv preprint arXiv:1706.02690, 2017. 2, 6

    arXiv 2017

  4. [3]

    Energy-based out-of- distribution detection,

    W. Liu, X. Wang, J. Owens, and Y . Li, “Energy-based out-of- distribution detection,”Advances in neural information pro- cessing systems, vol. 33, pp. 21 464–21 475, 2020. 4, 5, 6

    work page 2020

  5. [4]

    On the importance of gradi- ents for detecting distributional shifts in the wild,

    R. Huang, A. Geng, and Y . Li, “On the importance of gradi- ents for detecting distributional shifts in the wild,”Advances in Neural Information Processing Systems, vol. 34, pp. 677– 689, 2021. 1, 2, 6

    work page 2021

  6. [5]

    Vim: Out-of- distribution with virtual-logit matching,

    H. Wang, Z. Li, L. Feng, and W. Zhang, “Vim: Out-of- distribution with virtual-logit matching,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 4921–4930. 5, 6

    work page 2022

  7. [6]

    Out-of-distribution de- tection with deep nearest neighbors,

    Y . Sun, Y . Ming, X. Zhu, and Y . Li, “Out-of-distribution de- tection with deep nearest neighbors,” inInternational Con- ference on Machine Learning. PMLR, 2022, pp. 20 827– 20 840. 6

    work page 2022

  8. [7]

    V os: Learning what you don’t know by virtual outlier synthesis,

    X. Du, Z. Wang, M. Cai, and Y . Li, “V os: Learning what you don’t know by virtual outlier synthesis,”arXiv preprint arXiv:2202.01197, 2022. 6

    arXiv 2022

Show all 68 references
  1. [8]

    Dice: Leveraging sparsification for out- of-distribution detection,

    Y . Sun and Y . Li, “Dice: Leveraging sparsification for out- of-distribution detection,” inEuropean Conference on Com- puter Vision. Springer, 2022, pp. 691–708. 1, 6

    2022

  2. [9]

    React: Out-of-distribution detec- tion with rectified activations,

    Y . Sun, C. Guo, and Y . Li, “React: Out-of-distribution detec- tion with rectified activations,”Advances in Neural Informa- tion Processing Systems, vol. 34, pp. 144–157, 2021. 1, 2, 5, 6

    2021

  3. [10]

    Zero-shot out-of-distribution detection based on the pre-trained model clip,

    S. Esmaeilpour, B. Liu, E. Robertson, and L. Shu, “Zero-shot out-of-distribution detection based on the pre-trained model clip,” inProceedings of the AAAI conference on artificial in- telligence, vol. 36, no. 6, 2022, pp. 6568–6576. 5, 6

    2022

  4. [11]

    Delving into out-of-distribution detection with vision-language repre- sentations,

    Y . Ming, Z. Cai, J. Gu, Y . Sun, W. Li, and Y . Li, “Delving into out-of-distribution detection with vision-language repre- sentations,”Advances in neural information processing sys- tems, vol. 35, pp. 35 087–35 102, 2022. 1, 2, 6

    2022

  5. [12]

    Learning to prompt for vision-language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Learning to prompt for vision-language models,”International Journal of Com- puter Vision, vol. 130, no. 9, pp. 2337–2348, 2022. 6

    2022

  6. [13]

    Conditional prompt learning for vision-language models,

    ——, “Conditional prompt learning for vision-language models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 16 816– 16 825. 6

    2022

  7. [14]

    Non-parametric outlier synthesis,

    L. Tao, X. Du, X. Zhu, and Y . Li, “Non-parametric outlier synthesis,”arXiv preprint arXiv:2303.02966, 2023. 3, 6

    2023 arXiv

  8. [15]

    Clipn for zero-shot ood detection: Teaching clip to say no,

    H. Wang, Y . Li, H. Yao, and X. Li, “Clipn for zero-shot ood detection: Teaching clip to say no,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 1802–1812. 1, 2, 6

    2023

  9. [16]

    Out-of-distribution detection with negative prompts,

    J. Nie, Y . Zhang, Z. Fang, T. Liu, B. Han, and X. Tian, “Out-of-distribution detection with negative prompts,” in The Twelfth International Conference on Learning Represen- tations, 2024. 2, 6

    2024

  10. [17]

    Locoop: Few-shot out-of-distribution detection via prompt learning,

    A. Miyai, Q. Yu, G. Irie, and K. Aizawa, “Locoop: Few-shot out-of-distribution detection via prompt learning,”Advances in Neural Information Processing Systems, vol. 36, 2024. 1, 6

    2024

  11. [18]

    Negative label guided ood detection with pretrained vision-language models,

    X. Jiang, F. Liu, Z. Fang, H. Chen, T. Liu, F. Zheng, and B. Han, “Negative label guided ood detection with pretrained vision-language models,”arXiv preprint arXiv:2403.20078,

    arXiv

  12. [19]

    Hierarchical visual categories modeling: A joint representation learning and density estimation frame- work for out-of-distribution detection,

    J. Li, X. Zhou, P. Guo, Y . Sun, Y . Huang, W. Ge, and W. Zhang, “Hierarchical visual categories modeling: A joint representation learning and density estimation frame- work for out-of-distribution detection,” inProceedings of the IEEE/CVF International Conference on Computer ...

    2023

  13. [20]

    Tagood: A novel approach to out- of-distribution detection via vision-language representations and class center learning,

    J. Li, X. Zhou, K. Jiang, L. Hong, P. Guo, Z. Chen, W. Ge, and W. Zhang, “Tagood: A novel approach to out- of-distribution detection via vision-language representations and class center learning,”arXiv preprint arXiv:2408.15566,

    arXiv

  14. [21]

    Wordnet: a lexical database for english,

    G. A. Miller, “Wordnet: a lexical database for english,”Com- munications of the ACM, vol. 38, no. 11, pp. 39–41, 1995. 3

    1995

  15. [22]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learn- ing. PMLR, 2021, pp. 8748–8763. 1

    2021

  16. [23]

    Visual instruction tun- ing,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tun- ing,” 2023. 1

    2023

  17. [24]

    Qwen technical report,

    J. Bai, S. Bai, Y . Chu, Z. Cui, K. Dang, X. Deng, Y . Fan, W. Ge, Y . Han, F. Huang, B. Hui, L. Ji, M. Li, J. Lin, R. Lin, D. Liu, G. Liu, C. Lu, K. Lu, J. Ma, R. Men, X. Ren, X. Ren, C. Tan, S. Tan, J. Tu, P. Wang, S. Wang, W. Wang, S. Wu, B. Xu, J. Xu, A. Yang, H. Yang, J. ...

    2023 arXiv

  18. [25]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,” inInternational conference on ma- chine learning. PMLR, 2023, pp. 19 730–19 742. 1

    2023

  19. [26]

    Gpt-4 technical report,

    OpenAI, “Gpt-4 technical report,” 2023, accessed: 2023-10-

    2023

  20. [27]

    Available: https://cdn.openai.com/papers/gpt- 4.pdf 1

    [Online]. Available: https://cdn.openai.com/papers/gpt- 4.pdf 1

  21. [28]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Om- mer, “High-resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684– 10 695. 1, 5

    2022

  22. [29]

    SDXL: Improving latent diffusion models for high-resolution im- age synthesis,

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dock- horn, J. M ¨uller, J. Penna, and R. Rombach, “SDXL: Improving latent diffusion models for high-resolution im- age synthesis,” inThe Twelfth International Conference on Learning Representations, 2024. [Online]. Available: ...

    2024

  23. [30]

    Auto-encoding variational bayes,

    D. P. Kingma, “Auto-encoding variational bayes,”arXiv preprint arXiv:1312.6114, 2013. 3

    2013 arXiv

  24. [31]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE confer- ence on computer vision and pattern recognition, 2016, pp. 770–778. 4, 7

    2016

  25. [32]

    U-net: Con- volutional networks for biomedical image segmenta- tion,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Con- volutional networks for biomedical image segmenta- tion,” inMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III

    2015

  26. [33]

    Springer, 2015, pp. 234–241. 3

    2015

  27. [34]

    Content-based unrestricted adversarial at- tack,

    Z. Chen, B. Li, S. Wu, K. Jiang, S. Ding, and W. Zhang, “Content-based unrestricted adversarial at- tack,” inAdvances in Neural Information Process- ing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, ...

    2023

  28. [35]

    Mos: Towards scaling out-of- distribution detection for large semantic space,

    R. Huang and Y . Li, “Mos: Towards scaling out-of- distribution detection for large semantic space,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 8710–8719. 5

    2021

  29. [36]

    Imagenet large scale visual recognition challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Imagenet large scale visual recognition challenge,” International journal of computer vision, vol. 115, pp. 211– 252, 2015. 5

    2015

  30. [37]

    The inat- uralist species classification and detection dataset,

    G. Van Horn, O. Mac Aodha, Y . Song, Y . Cui, C. Sun, A. Shepard, H. Adam, P. Perona, and S. Belongie, “The inat- uralist species classification and detection dataset,” inPro- ceedings of the IEEE conference on computer vision and pat- tern recognition, 2018, pp. 8769–8778. 5

    2018

  31. [38]

    Sun database: Large-scale scene recognition from abbey to zoo,

    J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba, “Sun database: Large-scale scene recognition from abbey to zoo,” in2010 IEEE computer society conference on com- puter vision and pattern recognition. IEEE, 2010, pp. 3485–

    2010

  32. [39]

    Places: A 10 million image database for scene recognition,

    B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba, “Places: A 10 million image database for scene recognition,” IEEE transactions on pattern analysis and machine intelli- gence, vol. 40, no. 6, pp. 1452–1464, 2017. 5

    2017

  33. [40]

    Describing textures in the wild,

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi, “Describing textures in the wild,” inProceed- ings of the IEEE conference on computer vision and pattern recognition, 2014, pp. 3606–3613. 5

    2014

  34. [41]

    Openood: Benchmarking generalized out-of-distribution detection,

    J. Yang, P. Wang, D. Zou, Z. Zhou, K. Ding, W. Peng, H. Wang, G. Chen, B. Li, Y . Sunet al., “Openood: Benchmarking generalized out-of-distribution detection,” Advances in Neural Information Processing Systems, vol. 35, pp. 32 598–32 611, 2022. 5

    2022

  35. [42]

    Openood v1. 5: Enhanced benchmark for out-of-distribution detection,

    J. Zhang, J. Yang, P. Wang, H. Wang, Y . Lin, H. Zhang, Y . Sun, X. Du, K. Zhou, W. Zhanget al., “Openood v1. 5: Enhanced benchmark for out-of-distribution detection,” arXiv preprint arXiv:2306.09301, 2023. 5, 6

    2023 arXiv

  36. [43]

    Open-set recognition: A good closed-set classifier is all you need,

    S. Vaze, K. Han, A. Vedaldi, and A. Zisserman, “Open-set recognition: A good closed-set classifier is all you need,” inInternational Conference on Learning Representations,

  37. [44]

    In or out? fix- ing imagenet out-of-distribution detection evaluation,

    J. Bitterwolf, M. Mueller, and M. Hein, “In or out? fix- ing imagenet out-of-distribution detection evaluation,”arXiv preprint arXiv:2306.00826, 2023. 5

    2023 arXiv

  38. [45]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022. 1

    2021

  39. [46]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” 2021. [Online]. Available: https://arxi...

    2021 arXiv

  40. [47]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” inProceed- ings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 4700–4708

    2017

  41. [48]

    Aggre- gated residual transformations for deep neural networks,

    S. Xie, R. Girshick, P. Doll ´ar, Z. Tu, and K. He, “Aggre- gated residual transformations for deep neural networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1492–1500. 1

    2017

  42. [49]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al., “Pytorch: An imperative style, high-performance deep learning library,”Advances in neural information processing systems, vol. 32, 2019. 5

    2019

  43. [50]

    Gen: Pushing the limits of softmax-based out-of-distribution detection,

    X. Liu, Y . Lochman, and C. Zach, “Gen: Pushing the limits of softmax-based out-of-distribution detection,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 23 946–23 955. 6

    2023

  44. [51]

    Augmix: A simple data process- ing method to improve robustness and uncertainty,

    D. Hendrycks, N. Mu, E. D. Cubuk, B. Zoph, J. Gilmer, and B. Lakshminarayanan, “Augmix: A simple data process- ing method to improve robustness and uncertainty,”arXiv preprint arXiv:1912.02781, 2019. 6

    1912 arXiv

  45. [52]

    A simple fix to mahalanobis dis- tance for improving near-ood detection,

    J. Ren, S. Fort, J. Liu, A. G. Roy, S. Padhy, and B. Lakshminarayanan, “A simple fix to mahalanobis dis- tance for improving near-ood detection,”arXiv preprint arXiv:2106.09022, 2021. 6

    2021 arXiv

  46. [53]

    Extremely simple activation shaping for out-of-distribution detection,

    A. Djurisic, N. Bozanic, A. Ashok, and R. Liu, “Extremely simple activation shaping for out-of-distribution detection,” arXiv preprint arXiv:2209.09858, 2022. 6

    2022 arXiv

  47. [54]

    Natural adversarial examples,

    D. Hendrycks, K. Zhao, S. Basart, J. Steinhardt, and D. Song, “Natural adversarial examples,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 15 262–15 271. 5

    2021

  48. [55]

    A simple unified frame- work for detecting out-of-distribution samples and adversar- ial attacks,

    K. Lee, K. Lee, H. Lee, and J. Shin, “A simple unified frame- work for detecting out-of-distribution samples and adversar- ial attacks,”Advances in neural information processing sys- tems, vol. 31, 2018. 2

    2018

  49. [56]

    Contrastive training for improved out-of-distribution detection,

    J. Winkens, R. Bunel, A. G. Roy, R. Stanforth, V . Natarajan, J. R. Ledsam, P. MacWilliams, P. Kohli, A. Karthikesalingam, S. Kohlet al., “Contrastive training for improved out-of-distribution detection,”arXiv preprint arXiv:2007.05566, 2020. 2

    2007 arXiv

  50. [57]

    A boundary based out-of-distribution classifier for generalized zero-shot learn- ing,

    X. Chen, X. Lan, F. Sun, and N. Zheng, “A boundary based out-of-distribution classifier for generalized zero-shot learn- ing,” inEuropean conference on computer vision. Springer, 2020, pp. 572–588. 2

    2020

  51. [58]

    Out-of-distribution detection using union of 1-dimensional subspaces,

    A. Zaeemzadeh, N. Bisagno, Z. Sambugaro, N. Conci, N. Rahnavard, and M. Shah, “Out-of-distribution detection using union of 1-dimensional subspaces,” inProceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, 2021, pp. 9452–9461. 2

    2021

  52. [59]

    Un- certainty estimation using a single deep deterministic neural network,

    J. Van Amersfoort, L. Smith, Y . W. Teh, and Y . Gal, “Un- certainty estimation using a single deep deterministic neural network,” inInternational conference on machine learning. PMLR, 2020, pp. 9690–9700. 2

    2020

  53. [60]

    Feature space singularity for out-of-distribution detection,

    H. Huang, Z. Li, L. Wang, S. Chen, B. Dong, and X. Zhou, “Feature space singularity for out-of-distribution detection,” arXiv preprint arXiv:2011.14654, 2020. 2

    2011 arXiv

  54. [61]

    Exploring the limits of out-of-distribution detection,

    S. Fort, J. Ren, and B. Lakshminarayanan, “Exploring the limits of out-of-distribution detection,”Advances in Neural Information Processing Systems, vol. 34, pp. 7068–7081,

  55. [63]

    Lapt: Label-driven automated prompt tuning for ood detection with vision- language models,

    Y . Zhang, W. Zhu, C. He, and L. Zhang, “Lapt: Label-driven automated prompt tuning for ood detection with vision- language models,” inEuropean Conference on Computer Vi- sion. Springer, 2024, pp. 271–288. 3

    2024

  56. [64]

    Dream the impossi- ble: Outlier imagination with diffusion models,

    X. Du, Y . Sun, J. Zhu, and Y . Li, “Dream the impossi- ble: Outlier imagination with diffusion models,”Advances in Neural Information Processing Systems, vol. 36, pp. 60 878– 60 901, 2023. 3

    2023

  57. [65]

    Conjugated semantic pool im- proves ood detection with pre-trained vision-language mod- els,

    M. Chen, J. Gao, and C. Xu, “Conjugated semantic pool im- proves ood detection with pre-trained vision-language mod- els,”Advances in Neural Information Processing Systems, vol. 37, pp. 82 560–82 593, 2024. 6

    2024

  58. [66]

    Adaneg: Adaptive negative proxy guided OOD detection with vision-language models,

    Y . Zhang and L. Zhang, “Adaneg: Adaptive negative proxy guided OOD detection with vision-language models,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. [Online]. Available: https: //openreview.net/forum?id=vS5NC7jtCI 6

    2024

  59. [67]

    Is out- of-distribution detection learnable?

    Z. Fang, Y . Li, J. Lu, J. Dong, B. Han, and F. Liu, “Is out- of-distribution detection learnable?”Advances in Neural In- formation Processing Systems, vol. 35, pp. 37 199–37 213,

  60. [68]

    Out-of-distribution detection learning with unreli- able out-of-distribution sources,

    H. Zheng, Q. Wang, Z. Fang, X. Xia, F. Liu, T. Liu, and B. Han, “Out-of-distribution detection learning with unreli- able out-of-distribution sources,”Advances in neural infor- mation processing systems, vol. 36, pp. 72 110–72 123, 2023. 2

    2023

Pith tools

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