Pith. sign in

REVIEW 3 major objections 5 minor 46 references

Object-aware masks and staged decoupling let synthetic defects localize better than prior generators, reaching 88.1 AP-P on MVTec AD.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 05:07 UTC pith:UUMEWCD6

load-bearing objection Solid empirical paper on anomaly generation with a genuinely careful evaluation, but the load-bearing Stage 2 coarse-mask detail is parked in a missing appendix and the gains over the strongest baseline are modest and uneven across categories. the 3 major comments →

arxiv 2607.29533 v1 pith:UUMEWCD6 submitted 2026-07-31 cs.CV

OSAGEN: Object-Aware Mask Priors and Multistage Decoupled Diffusion for Industrial Anomaly Generation

Jinyi Xu , Peng Chen , Yunkang Cao , Chengliang Liu , Xinghui Dong , Chao Huang This is my paper
classification cs.CV
keywords industrial anomaly detectionanomaly generationdiffusion modelsmask-guided inpaintingfew-shot learningsynthetic dataobject-aware masksanomaly localization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

OSAGEN claims that the two biggest failures of mask-guided synthetic anomaly generation—defects that slavishly copy the condition mask's geometry, and mask priors that ignore the object they will be placed on—can be addressed together. It separates learning into three stages: normal appearance, defect appearance under a deliberately loose condition mask, and fine-grained mask calibration, so the generator treats the mask as a candidate region rather than a boundary. A second module, Query-Bias-Gen, injects structure from a matched normal image into a mask-diffusion model, producing masks that fit the current object instance. With these pieces, plus inference-time spatial control and a label-materialization step, the paper reports the best pixel-level localization among the compared generators on MVTec AD and VisA, with AP-P/F1-P of 88.1/82.2 and 68.5/66.1. A sympathetic reader would care because better synthetic pairs promise to shrink the real-annotation bottleneck in industrial visual inspection.

Core claim

The central claim is that decoupling defect learning from the condition mask's exact geometry is what unlocks both realism and accurate supervision. OSAGEN's three-stage curriculum iterates from random-mask normal learning, to defect learning under a coarse 'loose candidate' mask with supervision confined to the true defect, to final calibration on fine-grained masks. This staged design is what lets inference-time Spatial-CFG compare defect and normal branches without forcing mask-shaped anomalies. Query-Bias-Gen supplies the required object-aware support by biasing mask-diffusion queries with structure from a matched normal source image, and a materialization step recovers pixel-level label

What carries the argument

The load-bearing mechanism is the three-stage decoupled inpainting curriculum, specifically Stage 2's separation of the condition mask m_coarse (a loose candidate region) from the supervision mask m_gt (the true defect). Query-Bias-Gen is the complementary object-awareness mechanism: a query-bias injection into a mask latent diffusion model, so that the sampled mask prior is conditioned on the current object instance. ISC (Inference-time Spatial Control) transfers training-time constraints to sampling via spatial CFG and attention gating, and materialization converts the loose condition into pixel labels.

Load-bearing premise

Stage 2's decoupling relies on the coarse condition mask being a true loose superset of the defect with a sufficient margin of normal content; if the margin is too small, defect semantics re-couple to the condition geometry and the reported reductions in mask over-following would not hold.

What would settle it

Construct m_coarse as m_gt dilated by only one pixel on a category where OSAGEN currently works, retrain Stage 2 under the same budget, and measure AR-FR/S-FR and downstream AP-P. If the over-following diagnostic exceeds the variant without Stage 2, or AP-P drops below the w/o-Stage-2 ablation, the decoupling claim collapses. A simpler observer-level test: generate 1000 pairs and check whether realized defects stretch to fill the condition mask when the margin is tight.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Synthetic image–mask pairs from OSAGEN can train a downstream localization model to perform better on real anomalies than pairs from prior generators under the same protocol on both MVTec AD and VisA.
  • The staged decoupling reduces mask over-following, as measured by the AR-FR/S-FR diagnostics on enlarged condition masks.
  • Object-aware mask priors improve mask placement on the object: matched-source conditioning retains 87.5% Object Hit Rate under rotation, versus 50.1% for shuffled sources.
  • The full pipeline remains competitive in image-level detection while leading pixel-level localization, and it is cheaper per generated pair than several prior generators.
  • The materialization step provides a general label-alignment benefit even for other generators, but yields a smaller gain for OSAGEN, indicating better intrinsic alignment between its visible defects and spatial support.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the decoupling claim is right, mask-guided anomaly generation no longer needs precise masks at training time, which could loosen annotation requirements for synthetic data pipelines (e.g., coarse scribbles instead of pixel masks).
  • Query-bias conditioning suggests a general recipe: any mask-prior generator could be made object-aware by injecting source-image structure into cross-attention queries, potentially benefiting other conditional synthesis tasks beyond anomalies.
  • The staged curriculum also suggests a transferable schedule for personalizing diffusion models when appearance, condition, and control signals compete, which could apply to other few-shot generation settings.
  • A testable extension is applying OSAGEN to unseen objects without category-specific adaptation; the paper defers this to future work, but the object-aware mask prior and materialization are prime candidates for zero-shot generalization.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. OSAGEN is a three-stage diffusion-based framework for generating synthetic industrial anomaly image-mask pairs. Stage 1 learns normal appearance with random masks; Stage 2 learns defect semantics under a coarse condition mask while applying loss only inside the ground-truth mask; Stage 3 calibrates to fine masks. Query-Bias-Gen injects object structure from a matched normal image into a mask latent diffusion model via additive query bias. During inference, ISC combines Attn-Gate, Spatial-CFG, and Soft-Normal-Latent-Blend to restrict anomaly propagation and preserve background, and a materialization predictor converts the condition region into pixel-level labels. The authors report AP-P/F1-P of 88.1/82.2 on MVTec AD and 68.5/66.1 on VisA under a unified downstream localization protocol, outperforming AnomalyDiffusion, DualAnoDiff, SeaS, and O2MAG. Ablations and controlled diagnostics identify the contributions of each component.

Significance. If the results hold, the paper makes a useful empirical contribution: it proposes a concrete recipe for synthetic anomaly-image generation with pixel labels, and it evaluates under a unified downstream protocol with two benchmarks. The label-source diagnostic (Table 6), the object-awareness diagnostic (Figure 6, OHR), and the explicit comparison against training-free generation (O2MAG) are strengths. The claim is falsifiable and the protocol is shared across methods. However, validation is incomplete: several load-bearing details are deferred to supplementary appendices that are not present in the submitted version, and the main quantitative claims are reported without error bars. These issues prevent the contribution from being fully verified as submitted.

major comments (3)
  1. [Method, 'Staged Defect-Conditioned Inpainting' (Eq. 4)] The central Stage-2 decoupling claim hinges on m_coarse being a loose superset of m_gt. The main text states this but defers the actual construction ('grouped-min quad construction and its automatic feasibility fallback') to Appendix D, which is not included. Stage 2 contributes the largest stage-level gain in Table 5 (5.4 AP-P), and the AR-FR/S-FR over-following diagnostic is supposed to demonstrate the mechanism. The revision must specify the construction, report the resulting margin distribution (e.g., area ratio m_coarse/m_gt), and include a sensitivity analysis with tight versus loose margins. Without this, the central mechanism is not verifiable.
  2. [Tables 2, 3, 5, 6; Appendix H/I] No standard deviations, confidence intervals, or seed counts are reported. The claimed advantages over the next best baseline are modest (MVTec AP-P 88.1 vs 86.3, F1-P 82.2 vs 80.8; VisA AP-P 68.5 vs 67.2), and the pipeline has multiple tunable strengths (omega_t, lambda_t, alpha_in, alpha_out, materialization threshold) whose sensitivity is only referenced to absent appendices G/H. Single-run results do not establish that the reported differences are outside run-to-run noise. Please report mean +/- std over at least three seeds for the main tables and key ablations, and include the sensitivity analysis in the main text or an available appendix.
  3. [Method, Query-Bias-Gen (Eq. 5-7) and Label Materialization] Several reproducibility-critical components are underspecified. The frozen visual encoder in Query-Bias-Gen is never named; the criterion for a 'structurally matched normal source image' is not defined; and the materialization predictor, which produces the final pixel labels, is described only at a high level with details deferred to Appendix K. The threshold selection and training-set construction for the predictor are essential to interpreting the reported downstream localization numbers. The revision should either include the full supplementary material or move these details into the main text.
minor comments (5)
  1. [Table 1 / Section 'Staged Defect-Conditioned Inpainting'] The notation m_c^s, m_l^s is used in Table 1 before it is defined in the text; define it before the table or restructure the table caption.
  2. [Eq. (8)] The values alpha_in and alpha_out are referenced to Appendix G; give default values or a brief indication of their magnitude so the reader can understand the gate behavior.
  3. [Table 4] KID values are reported without stating the unit or the number of samples used for the estimate; this makes cross-method comparison hard to interpret. Please state the evaluation protocol.
  4. [Related Work / References] The citation 'Chen, Huang, and Huang 2026' appears to be a self-citation in the introduction; if so, the authors should ensure it is necessary and clearly identifiable.
  5. [Figure 6] The acronym SAM3 is used without expansion; please define it on first use.

Circularity Check

0 steps flagged

No significant circularity: central results are externally benchmarked, with only non-load-bearing self-citations.

full rationale

This is a self-contained empirical claim rather than a derivation from definitions. OSAGEN's stated contributions (staged inpainting, Query-Bias-Gen, ISC, materialization) are evaluated by training a fixed downstream inspection model on generated pairs and measuring localization on held-out real anomalies from MVTec AD and VisA under the AnomalyDiffusion protocol. None of the method's own fitted parameters (LoRA modules, query-bias mapping layer, materialization predictor) appears in the evaluation metric; the materialization predictor is trained only on the permitted training split, and the label-source diagnostic explicitly separates label construction from image quality. Stage 2's decoupling claim relies on the coarse mask m_coarse being a loose superset of m_gt, and the construction is deferred to Appendix D, which is an omitted-support/verifiability concern, not a circularity: the ablation comparing w/o Stage 2 is an external comparison, and no equation reduces the reported AP-P/F1-P to a fitted input. The only self-citations (Chen, Huang, and Huang 2026; Sun et al. 2025; Yu et al. 2024) support background claims and are not load-bearing. Score 2 reflects these minor non-load-bearing self-citations, not actual circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper does not introduce new physical entities. Its central claim depends on several domain assumptions that are empirically supported within the paper but not proven: the utility of synthetic pairs, the transferability of a frozen visual encoder representation to mask diffusion, the validity of the Stage-1 branch as a normal baseline, and the transfer of the materialization predictor from real anomalies to generated images. The main free parameters are the inference-time guidance strengths and timestep schedules of ISC, which are dataset-level choices deferred to appendices.

free parameters (4)
  • ISC guidance strengths (ω_t, λ_t, α_in, α_out) = Not specified (Appendix G)
    Dataset-level guidance strengths for Spatial-CFG and Soft-Normal-Latent-Blend are hand-tuned per dataset; these directly control how much defect amplification and background blending occurs, and the performance depends on them.
  • Timestep schedules T_cfg and T_blend = Not specified (Appendix G)
    The sets of timesteps where Spatial-CFG and blending are active are chosen by hand; they constrain the inference-time behavior.
  • Materialization threshold = Not specified
    The predictor's anomaly probability threshold for converting probabilities to pixel labels is a decision variable; no automatic selection is described in the main text.
  • Stage-2 coarse mask construction parameters = Not specified (Appendix D)
    The coarse condition mask m_coarse is constructed via a 'grouped-min quad construction' with an automatic feasibility fallback, which likely involves size/geometry parameters that affect the decoupling.
axioms (4)
  • domain assumption Synthetic anomaly image-mask pairs improve downstream anomaly localization when used as training data
    The paper's entire contribution assumes that the generated pairs, when used to train a downstream model, improve real anomaly localization. This is supported by prior work but not re-derived.
  • domain assumption The frozen visual encoder provides object-structure representation aligned with the mask LDM's query space after a lightweight mapping
    Query-Bias-Gen relies on a frozen visual encoder (unnamed) and a learnable mapping to inject structure; the paper does not prove this alignment is sufficient, only demonstrates it empirically on six categories.
  • domain assumption The Stage-1 normal branch trained with random masks provides a valid 'normal baseline' for Spatial-CFG
    ISC substitutes the normal branch's noise prediction as the baseline in Eq. (9); this assumes the normal branch outputs a clean normal-image-like prediction for any inference mask.
  • domain assumption DINOv2 features plus pixel differences are sufficient to predict anomaly labels for generated images
    Materialization uses a frozen DINOv2 backbone and pixel differences; this assumes the predictor trained on real anomalies with degraded masks transfers to generated images.

pith-pipeline@v1.3.0-daily-deepseek · 12396 in / 14018 out tokens · 188157 ms · 2026-08-03T05:07:39.721677+00:00 · methodology

0 comments
read the original abstract

Industrial anomaly detection and localization are limited by scarce real anomalies and pixel-level annotations, a bottleneck that synthetic image-mask pairs can alleviate. However, existing few-shot mask-guided generation may over-follow mask geometry, produce weak anomalies, or use condition masks incompatible with the current object instance. We propose OSAGEN, which combines object-aware mask priors with multistage decoupled diffusion. Its three-stage adaptation sequentially learns normal appearance, defect appearance under coarse conditions, and fine-grained mask calibration, improving defect realization and local control. QBG injects object structure from a matched normal image into mask diffusion to produce object-aware priors, while ISC restricts anomaly propagation and preserves normal content during sampling. A lightweight materialization step recovers pixel-level labels aligned with the realized defects. On MVTec AD and VisA, OSAGEN achieves AP-P/F1-P scores of 88.1/82.2 and 68.5/66.1, respectively, under a unified downstream localization protocol. The code will be released upon acceptance.

Figures

Figures reproduced from arXiv: 2607.29533 by Chao Huang, Chengliang Liu, Jinyi Xu, Peng Chen, Xinghui Dong, Yunkang Cao.

Figure 1
Figure 1. Figure 1: Motivation of OSAGEN. Top: single-stage adaptation jointly learns normal appearance, defect semantics, and mask [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: OSAGEN overall framework with three components. Staged inpainting (top) learns normal appearance, defect [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual analysis of the staged inpainting curriculum. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of mask placement with [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on representative MVTec [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 3 linked inside Pith

  1. [1]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  2. [2]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Dual-interrelated diffusion model for few-shot anomaly image generation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  3. [3]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Defectfill: Realistic defect generation with inpainting diffusion model for visual inspection , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  4. [4]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    One-to-More: High-Fidelity Training-Free Anomaly Generation with Attention Control , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  5. [5]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Seas: Few-shot industrial anomaly image generation with separation and sharing fine-tuning , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  6. [6]

    arXiv preprint arXiv:2507.02314 , year =

    Choi, JaeHyuck and Kim, MinJun and Hong, Je Hyeong , title =. arXiv preprint arXiv:2507.02314 , year =

    Pith/arXiv arXiv

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    CutPaste: Self-Supervised Learning for Anomaly Detection and Localization , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  8. [8]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    DRAEM--A Discriminatively Trained Reconstruction Embedding for Surface Anomaly Detection , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  9. [9]

    Proceedings of the AAAI Conference on Artificial Intelligence , pages =

    Few-Shot Defect Image Generation via Defect-Aware Feature Manipulation , author =. Proceedings of the AAAI Conference on Artificial Intelligence , pages =

  10. [10]

    International Conference on Learning Representations , year =

    An Image is Worth One Word: Personalizing Text-to-Image Generation Using Textual Inversion , author =. International Conference on Learning Representations , year =

  11. [11]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  12. [12]

    Proceedings of the AAAI Conference on Artificial Intelligence , pages=

    Anomagic: Crossmodal prompt-driven zero-shot anomaly generation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=

  13. [13]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    High-Resolution Image Synthesis with Latent Diffusion Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  14. [14]

    Transactions on Machine Learning Research , year =

    DINOv2: Learning Robust Visual Features without Supervision , author =. Transactions on Machine Learning Research , year =

  15. [15]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    MVTec AD--A comprehensive real-world dataset for unsupervised anomaly detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  16. [16]

    European conference on computer vision , pages=

    Spot-the-difference self-supervised pre-training for anomaly detection and segmentation , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  17. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Reasoning-Driven Anomaly Detection and Localization with Image-Level Supervision , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  18. [18]

    Proceedings of the AAAI Conference on Artificial Intelligence , pages=

    MaskAD: Parallel Masked Autoencoder for Multi-class Unsupervised Anomaly Detection , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=

  19. [19]

    Advances in Neural Information Processing Systems , volume=

    Normal-abnormal guided generalist anomaly detection , author=. Advances in Neural Information Processing Systems , volume=

  20. [20]

    Pattern Recognition , pages=

    DyC-CLIP: Dynamic Context-Aware Multi-Modal Prompt Learning for Zero-Shot Anomaly Detection , author=. Pattern Recognition , pages=. 2026 , publisher=

    2026

  21. [21]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Omni-AD: A Large-scale and Versatile Benchmark for Industrial Anomaly Detection , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  22. [22]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    GPFlow: Gaussian Prototype Probability Flow for Unsupervised Multi-Modal Anomaly Detection , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  23. [23]

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

    Generate Aligned Anomaly: Region-Guided Few-Shot Anomaly Image-Mask Pair Synthesis for Industrial Inspection , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  24. [24]

    Proceedings of the AAAI Conference on Artificial Intelligence , pages=

    Quality-aware language-conditioned local auto-regressive anomaly synthesis and detection , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=

  25. [25]

    IEEE Access , year=

    Anodapter: a unified framework for generating aligned anomaly images and masks using diffusion models , author=. IEEE Access , year=

  26. [26]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    ReMP-AD: Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  27. [27]

    Natural Synthetic Anomalies for Self-Supervised Anomaly Detection and Localization , booktitle =

    Schl. Natural Synthetic Anomalies for Self-Supervised Anomaly Detection and Localization , booktitle =

  28. [28]

    European Conference on Computer Vision , pages =

    Chen, Qiyu and Luo, Huiyuan and Lv, Chengkan and Zhang, Zhengtao , title =. European Conference on Computer Vision , pages =

  29. [29]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Zhang, Ximiao and Xu, Min and Zhou, Xiuzhuang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  30. [30]

    IEEE Transactions on Automation Science and Engineering , volume =

    Niu, Shuanlong and Li, Bin and Wang, Xinggang and Lin, Hui , title =. IEEE Transactions on Automation Science and Engineering , volume =

  31. [31]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages =

    Zhang, Gongjie and Cui, Kaiwen and Hung, Tzu-Yi and Lu, Shijian , title =. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages =

  32. [32]

    European Conference on Computer Vision , pages =

    Gui, Guan and Gao, Bin-Bin and Liu, Jun and Wang, Chengjie and Wu, Yunsheng , title =. European Conference on Computer Vision , pages =

  33. [33]

    IEEE Access , volume=

    AnomalyControl: few-shot anomaly generation by controlnet inpainting , author=. IEEE Access , volume=. 2024 , publisher=

    2024

  34. [34]

    arXiv preprint arXiv:2404.19444 , year =

    Hu, Jie and Huang, Yawen and Lu, Yilin and Xie, Guoyang and Jiang, Guannan and Zheng, Yefeng and Lu, Zhichao , title =. arXiv preprint arXiv:2404.19444 , year =

    Pith/arXiv arXiv

  35. [35]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Unseen Visual Anomaly Generation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  36. [36]

    IEEE Transactions on Circuits and Systems for Video Technology , volume=

    TF ^2 : Few-Shot Text-Free Training-Free Defect Image Generation for Industrial Anomaly Inspection , author=. IEEE Transactions on Circuits and Systems for Video Technology , volume=. 2024 , publisher=

    2024

  37. [37]

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

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

  38. [38]

    International Conference on Learning Representations , year=

    Prompt-to-Prompt Image Editing with Cross-Attention Control , author=. International Conference on Learning Representations , year=

  39. [39]

    ACM transactions on Graphics (TOG) , volume=

    Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models , author=. ACM transactions on Graphics (TOG) , volume=. 2023 , publisher=

    2023

  40. [40]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  41. [41]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Plug-and-play diffusion features for text-driven image-to-image translation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  42. [42]

    Demystifying

    Bi. Demystifying. International Conference on Learning Representations , year =

  43. [43]

    and Lee, Yong Jae and Shechtman, Eli and Zhang, Richard , title =

    Ojha, Utkarsh and Li, Yijun and Lu, Jingwan and Efros, Alexei A. and Lee, Yong Jae and Shechtman, Eli and Zhang, Richard , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  44. [44]

    Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages=

    Free lunch of image-mask alignment for anomaly image generation and segmentation , author=. Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages=

  45. [45]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

    He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

  46. [46]

    arXiv preprint arXiv:2511.16719 , year =

    Carion, Nicolas and Gustafson, Laura and Hu, Yuan-Ting and Debnath, Shoubhik and Hu, Ronghang and Suris, Didac and Ryali, Chaitanya and Alwala, Kalyan Vasudev and Khedr, Haitham and Huang, Andrew and Lei, Jie and Ma, Tengyu and Guo, Baishan and Kalla, Arpit and Marks, Markus and Greer, Joseph and Wang, Meng and Sun, Peize and R. arXiv preprint arXiv:2511....

    Pith/arXiv arXiv