Pith. sign in

REVIEW 2 major objections 6 minor 146 references

Continual AI-generated image detectors forget through two coupled paths — feature drift and decision-boundary drift — and a decoupled fix for both reaches 99.36% average accuracy with 0.39% forgetting.

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

2026-07-31 23:42 UTC pith:ZK7777DA

load-bearing objection A well-specified continual AIGC detection method with a plausible dual-degradation story and strong reported numbers, but the central evidence for the decision-drift mechanism is thinner than the headline claims and the exemplar-memory premise is unvalidated. the 2 major comments →

arxiv 2607.27882 v1 pith:ZK7777DA submitted 2026-07-30 cs.CV

DECODE: Tackling Representation and Decision Degradation in Continual AI-Generated Image Detection

Zihao Cai , Xinghan Li , Ruiyan Yang , Xue Song , Haijun Shan , Jingjing Chen This is my paper
classification cs.CV
keywords continual learningAI-generated image detectioncatastrophic forgettingrepresentation degradationdecision boundary driftlow-rank adaptationridge regressionDual Degradation
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.

The paper argues that catastrophic forgetting in continually trained AI-generated image detectors is not only a representation problem. Even when the feature representation stays discriminative, the classifier's decision boundary can drift as the classification head is re-optimized on new generator domains. The authors call this compound failure Dual Degradation and show that fixing only the feature side leaves substantial accuracy on the table. Their DECODE framework attacks both paths: Subspace Diversity Regularization keeps LoRA adapter updates diverse and balanced, and Closed-Form Decision Alignment re-solves a ridge-regression classifier on stored exemplars after each adapter merge. If the reported numbers hold, DECODE reaches 99.36% average accuracy with 0.39% forgetting over eight sequential tasks and 95.36% accuracy on eleven generators never seen in training.

Core claim

The central claim is that representation-level stabilization alone is incomplete for continual AI-generated image detection. The paper demonstrates empirically that sequentially optimizing a shared classifier head on new domains shifts the decision boundary, so a detector can have high feature separability but still fail on earlier tasks. The proposed remedy is to treat representation and decision degradation as separate problems solved at separate stages: during adapter training, regularize the low-rank update geometry; after each merge, recalibrate the head with a closed-form ridge regression, with leave-one-out cross-validation selecting the regularization strength automatically. The auth

What carries the argument

The two load-bearing components are SDR and CDA. SDR operates on the LoRA factor matrices A and B for each task: an orthogonality loss penalizes off-diagonal entries of Gram matrices built from stacked A and B factors, encouraging intra-task directional diversity and inter-task subspace separation, while an energy-balance loss penalizes the variance of squared factor norms across rank dimensions to prevent a few rank-one directions from dominating. CDA is the decision-side machinery: after each adapter is merged into the backbone, features of stored exemplars are re-extracted, and the classification head is recomputed in closed form as a ridge regression solution; the regularization strength

Load-bearing premise

The whole decision-alignment gain rests on the assumption that 50 randomly stored exemplars per class, re-encoded after each merge, adequately represent each learned generator domain, so the ridge-regression head and its automatically chosen λ align the boundary for the full population; the paper reports no sensitivity analysis over m or sampling variance.

What would settle it

Run the same eight-task Protocol 1 with exemplar memory sizes m = 5, 10, 20, 50, 100 per class (or with repeated random draws at m=50) and compare final average accuracy and forgetting. If average accuracy drops toward the strongest replay baseline (91.30%) when m is small or varies widely across draws, the coverage premise fails. A second check: measure the agreement between the ridge-regression head's decisions on stored exemplars and on a much larger held-out test set per task; large disagreement would indicate the 50-exemplar memory is not representative.

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

If this is right

  • If Dual Degradation is the right diagnosis, then any continual detection method that only preserves features will hit a ceiling; decision-boundary recalibration is needed to recover the remaining accuracy.
  • The closed-form head recalibration procedure removes manual hyperparameter tuning for the classifier stage, making the method applicable to arbitrary-length task sequences without per-task validation.
  • Because the adapter is merged into a single backbone, DECODE keeps inference cost and test-time task identification independent of the number of seen generators — a property that matters for deployment.
  • The strong open-world numbers (95.36% on 11 unseen generators) suggest the representation-side regularization, especially SDR, transfers forensic cues across synthesis paradigms, not just retains them.
  • Under common perturbations (JPEG compression, resizing, Gaussian blur), the decoupled design maintains an edge over baselines, suggesting the dual-degradation fix is not brittle to mild post-processing.

Where Pith is reading between the lines

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

  • A testable extension is to vary the exemplar memory size m and task order; the paper reports only m=50 and one primary order. If the margin over the strongest replay baseline persists at m≈10, the CDA mechanism is robust; if it collapses, the headline numbers depend on a specific memory budget.
  • The Dual Degradation diagnosis likely transfers to other continual binary-forensics tasks (deepfake audio, steganalysis, synthetic-text detection), where subtle domain-specific traces and a continually re-optimized head create the same two-pathway failure.
  • Because CDA re-fits a linear head on re-encoded exemplars, it could be composed with any future representation-stabilizing method, suggesting that decision-level alignment is an orthogonal plug-in rather than a rival approach.
  • The paper's explanation of open-world gains leans on SDR, but does not isolate which of the three SDR terms (intra-task diversity, inter-task separation, energy balance) drives transfer; a targeted ablation on unseen generators would clarify the mechanism.

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

2 major / 6 minor

Summary. This paper studies continual learning for AI-generated image detection. It argues that forgetting in this setting cannot be explained by representation-level degradation alone; the decision boundary also drifts as the classification head is optimized on new domains, and both effects together constitute "Dual Degradation." The proposed DECODE has two components: Subspace Diversity Regularization (SDR), which penalizes intra-task directional redundancy, inter-task subspace overlap, and rank-energy imbalance in LoRA adapters; and Closed-Form Decision Alignment (CDA), which after each adapter merge re-estimates the classification head by ridge regression on features of exemplar memories, with λ selected by leave-one-out cross-validation. Experiments on 8 sequential generators report 99.36% average accuracy and 0.39% forgetting, and 95.36% average accuracy on 11 unseen generators. The paper includes ablations, a comparison to gradient-descent head recalibration, and robustness to JPEG, resizing, and blur.

Significance. If the results are reproducible, the paper makes a useful conceptual point: even with preserved features, the head can be misaligned, and a closed-form recalibration with automatic regularization is sufficient. The mathematical presentation is clean: Eq. (4) is the standard ridge solution with unpenalized bias, and Eq. (5) is the correct LOO/leverage computation. The design is well matched to the claim, with SDR-only and CDA-only ablations and a GD control. The open-world evaluation is broad (19 training + 11 unseen generators). The main uncertainty is whether the small margins attributable to CDA are robust to exemplar sampling; the paper does not currently provide this evidence.

major comments (2)
  1. [Closed-Form Decision Alignment, Exemplar Memory] The CDA mechanism depends entirely on the premise that m=50 random exemplars per class "provide sufficient coverage of the input distribution." This premise is asserted, not demonstrated. No sensitivity analysis over m, no repeated-sampling stability diagnostics, and no coverage measures are provided. Table 3 shows the incremental effect of CDA over SDR is 0.22 percentage points (99.10→99.36 AA) and over GD is 0.25 (99.11→99.36); without an estimate of variance over exemplar draws, these margins are within the range that could arise from a favorable draw. I request (i) accuracy/forgetting for at least three different m values (e.g., 10, 50, 100, 200), (ii) repeated exemplar draws with standard deviations, and (iii) a simple coverage diagnostic (e.g., comparison of exemplar feature distribution to the full task distribution).
  2. [Tables 1 and 3] The paper states each experiment is repeated three times and mean results are reported, but no standard deviations or confidence intervals are provided for any of the headline numbers. For the central comparison in Table 3, CDA vs GD differs by 0.25 AA and 0.41 AF at the end of the sequence; Table 1's DECODE vs Tang et al. replay baseline differs by 8.06 AA. The former is the evidence for the decision-degradation mechanism, and without variance it is not established. Please report error bars and, if possible, paired differences over seeds.
minor comments (6)
  1. [Open-World Generalization Performance] The claim "DECODE performs best on every generator" is not accompanied by measures of variance; the margins on GPT-Image-1.5 (92.00 vs 90.91) and Seedream4.5 (92.50 vs 91.27) are small. Please add per-generator variance or significance statements.
  2. [Supplementary material] The text refers to the supplementary material for additional empirical evidence and ablation results, but no supplementary is included in the submission. Please include it.
  3. [Figure 1] The decision-boundary panel is described qualitatively. Please provide the underlying accuracy/AUC values and the number of tasks in the caption or figure.
  4. [Eq. (5)] The LOO-CV formula would benefit from a derivation or reference; the 1/N term for the unregularized bias is not obvious to a general reader.
  5. [Datasets and Implementation Details] The number of training images per task and per class is not specified. This is needed for reproducibility and to understand the effect of memory size in CDA.
  6. [General typesetting] Many spaces are missing between words in the abstract and body (e.g., "withonly0.39%forgetting"). Please ensure proper typesetting in the final version.

Circularity Check

0 steps flagged

No significant circularity: CDA is trained on exemplar memory and evaluated on held-out test sets, λ is selected by LOO-CV on the memory rather than test labels, and the central Dual Degradation claim is supported by independent ablations and external comparisons.

full rationale

I walked the claimed derivation chain and found no step where a prediction reduces by construction to a fitted input or to a self-citation. The closed-form decision head (Eq. 4) is trained on re-encoded exemplar features, and the regularization strength λ is selected by LOO-CV on the same exemplar memory (Eq. 5); however, the reported accuracies and forgetting metrics (Tables 1–2, 3) are evaluated on held-out test sets, so the fit is not the prediction. The paper also discloses that CDA is replay-based and ablates it against gradient-descent head updates using the identical exemplar memory, which controls for the trivial replay effect. SDR’s losses (Eqs. 1–3) are regularizers on LoRA factor geometry; they do not encode the evaluation metrics or the test-domain labels. No uniqueness theorem is imported from the authors’ own prior work, and the paper does not rely on self-citations to justify its central claim. The unvalidated assertion that m=50 random exemplars provide sufficient coverage is a robustness/generalization concern, not a circularity: even if that premise failed, the method would still be a head trained on memory and tested on separate data. Thus, no circular step is identifiable under the specified standards.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claim (dual degradation + DECODE's 99.36%/0.39%/95.36%) rests on: standard ridge/LOO math (correct), the CLIP-features-are-forensic premise, the exemplar-coverage premise, the merge-safety premise, and the SDR-diversity-means-forensic-diversity premise. Free parameters are r=8, m=50, λ_orth=λ_bal=1, α (unstated), and the training schedule; the LOO-CV λ grid is also unspecified. No invented entities: 'Dual Degradation' is a name for a proposed failure-mode decomposition, not a new entity with an independent falsifiable handle.

free parameters (5)
  • λ_orth = λ_bal = 1 = 1.0, 1.0
    Regularization weights in Eq. (3), hand-set for all tasks; no sensitivity analysis reported. They directly control SDR, which the paper credits for the 95.36% OOD performance.
  • LoRA rank r = 8
    Adapter subspace rank; capacity bound on SDR diversity. Set, not swept.
  • Exemplar memory size m = 50 per class
    Buffer size for CDA recalibration and λ selection; the paper's strongest replay-baseline gap is attributed to CDA, but m is never varied.
  • LoRA scaling α = not stated
    Standard LoRA scaling factor in ΔW = (α/r)BA; value not given in text, needed for exact replication. The LOO-CV candidate grid over λ is also unspecified.
  • Training schedule = LR 1e-3, batch 16, 10 epochs/task
    Hand-set training hyperparameters across all tasks; routine choices but unvalidated by sensitivity analysis.
axioms (6)
  • domain assumption Frozen CLIP ViT-L/14 features contain transferable real-vs-generated forensic traces across diffusion, GAN, face-manipulation, 3DGS, and autoregressive generators.
    Used implicitly throughout (Implementation Details: CLIP ViT-L/14 backbone). Load-bearing for the OOD claim (Table 2): if the pretrained encoder lacked cross-generator forensic signal, SDR's diversity preservation could not yield 95.36% on 11 unseen generators. The cited literature uses the same assumption, so it is a shared domain premise rather than an ad-hoc invention.
  • standard math Closed-form ridge solution (Eq. 4) and LOO-CV hat-matrix/leverage formulas (Eq. 5) are correct for the unpenalized-bias case.
    The 1/N term in H_ii and the e_i = (y − ŷ)/(1 − H_ii) LOO residual are the standard linear-smoother results; the derivation is not shown but the formulas check out.
  • domain assumption m=50 randomly sampled images per class, re-encoded by the merged backbone, provide sufficient coverage of each domain for CDA head recalibration.
    Stated in the Exemplar Memory paragraph: 'random sampling provides sufficient coverage of the input distribution.' No sensitivity analysis over m, no coverage diagnostics; the full DECODE advantage over the replay baseline (99.36 vs 91.30 AA) rests on this premise.
  • domain assumption Merging adapters into a single backbone W^{(k)} = W^{(k-1)} + (α/r)BA preserves learned forensic structure without task routing.
    Parameter Merging paragraph. SDR's orthogonality penalties are meant to make merges safe; the assumption that the merged weight is a valid single-model detector at every step is empirically validated by the reported sequence, not proven.
  • ad hoc to paper Intra-task rank-one directional diversity and balanced rank energy correspond to diverse, complementary forensic cues.
    The SDR design premise, motivated solely by Figure 3's geometry (energy concentration and Frobenius cosine redundancy in vanilla LoRA). The paper itself disclaims any one-to-one mapping between rank directions and artifact types, so the link between regularizer geometry and forensic-cue diversity is a bet supported only by the ablation.
  • domain assumption Protocol 1's single task ordering is representative for measuring continual detection performance.
    The order is 'deliberately designed as a heterogeneous stress test'; additional orders are only in the supplementary. Headline AA/AF numbers (99.36%/0.39%) are for this one order.

pith-pipeline@v1.3.0-daily-deepseek · 13743 in / 26746 out tokens · 261519 ms · 2026-07-31T23:42:57.135057+00:00 · methodology

0 comments
read the original abstract

As generative models continue to evolve, AI-generated image detectors must incrementally adapt to emerging generative domains while preserving knowledge acquired from previous ones. This continual learning setting is particularly challenging because forensic traces are often subtle and generator-specific, making detectors highly vulnerable to catastrophic forgetting. Existing methods primarily address this problem by stabilizing feature representations, implicitly treating forgetting as a representation-level issue. In this paper, we show that this perspective is incomplete. We demonstrate that even when feature representations remain discriminative, the decision boundary can progressively drift as the classification head is continually optimized on new domains. These two effects jointly give rise to a compound failure mode, termed Dual Degradation. To overcome this challenge, we propose DECODE, a decoupled continual detection framework that jointly mitigates representation- and decision-level forgetting. Specifically, we introduce Subspace Diversity Regularization (SDR) to preserve diverse forensic representations and Closed-Form Decision Alignment (CDA) to recalibrate the shared classification head after each adapter merge without manual hyperparameter tuning. Extensive experiments on 19 generative domains show that DECODE achieves an average accuracy of 99.36% with only 0.39% forgetting, while further generalizing to 11 unseen generators with 95.36% accuracy.

Figures

Figures reproduced from arXiv: 2607.27882 by Haijun Shan, Jingjing Chen, Ruiyan Yang, Xinghan Li, Xue Song, Zihao Cai.

Figure 1
Figure 1. Figure 1: Empirical evidence of Dual Degradation. (a) Feature drift on StyleGAN3 as new tasks are learned. Vanilla LoRA [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of DECODE. In Stage 1, SDR regularizes the low-rank adapter to preserve diverse forensic [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Energy concentration and directional redundancy [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Robustness under JPEG compression, resizing, and Gaussian blur on the Protocol 1 test sets. [PITH_FULL_IMAGE:figures/full_fig_p007_4.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

146 extracted references · 6 linked inside Pith

  1. [1]

    for now , author=

    CNN-generated images are surprisingly easy to spot... for now , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  2. [2]

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

    Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  3. [3]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  4. [4]

    arXiv preprint arXiv:2406.19435 , year=

    A sanity check for ai-generated image detection , author=. arXiv preprint arXiv:2406.19435 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2603.10598 , year=

    Layer Consistency Matters: Elegant Latent Transition Discrepancy for Generalizable Synthetic Image Detection , author=. arXiv preprint arXiv:2603.10598 , year=

    arXiv

  6. [6]

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

    Dire for diffusion-generated image detection , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  7. [7]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Improving synthetic image detection towards generalization: An image transformation perspective , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 , pages=

  8. [8]

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

    A bias-free training paradigm for more general ai-generated image detection , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  9. [9]

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

    Towards universal fake image detectors that generalize across generative models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  10. [10]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    C2p-clip: Injecting category common prompt in clip to enhance generalization in deepfake detection , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  11. [11]

    arXiv preprint arXiv:2508.01603 , year=

    Towards Generalizable AI-Generated Image Detection via Image-Adaptive Prompt Learning , author=. arXiv preprint arXiv:2508.01603 , year=

    arXiv

  12. [12]

    arXiv preprint arXiv:2411.15633 , year=

    Orthogonal subspace decomposition for generalizable ai-generated image detection , author=. arXiv preprint arXiv:2411.15633 , year=

    Pith/arXiv arXiv

  13. [13]

    Proceedings of the 29th ACM International Conference on Multimedia , pages=

    Cored: Generalizing fake media detection with continual representation using distillation , author=. Proceedings of the 29th ACM International Conference on Multimedia , pages=

  14. [14]

    Proceedings of the 31st ACM International Conference on Multimedia , pages=

    Dfil: Deepfake incremental learning by exploiting domain-invariant forgery clues , author=. Proceedings of the 31st ACM International Conference on Multimedia , pages=

  15. [15]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages=

    Dynamic mixed-prototype model for incremental deepfake detection , author=. Proceedings of the 32nd ACM International Conference on Multimedia , pages=

  16. [16]

    The Fourteenth International Conference on Learning Representations , year=

    HSIC Bottleneck for Cross-Generator and Domain-Incremental Synthetic Image Detection , author=. The Fourteenth International Conference on Learning Representations , year=

  17. [17]

    IEEE Transactions on Information Forensics and Security , year=

    Towards extensible detection of AI-generated images via content-agnostic adapter-based category-aware incremental learning , author=. IEEE Transactions on Information Forensics and Security , year=

  18. [18]

    arXiv preprint arXiv:2509.19230 , year=

    DevFD: Developmental Face Forgery Detection by Learning Shared and Orthogonal LoRA Subspaces , author=. arXiv preprint arXiv:2509.19230 , year=

    arXiv

  19. [19]

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

    Stacking brick by brick: Aligned feature isolation for incremental face forgery detection , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  20. [20]

    IEEE Transactions on Multimedia , year=

    Generalizable and Adaptive Continual Learning Framework for AI-generated Image Detection , author=. IEEE Transactions on Multimedia , year=

  21. [21]

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

    Generalization-Preserved Learning: Closing the Backdoor to Catastrophic Forgetting in Continual Deepfake Detection , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  22. [22]

    arXiv preprint arXiv:2512.00539 , year=

    SAIDO: Generalizable Detection of AI-Generated Images via Scene-Aware and Importance-Guided Dynamic Optimization in Continual Learning , author=. arXiv preprint arXiv:2512.00539 , year=

    arXiv

  23. [23]

    Proceedings of the national academy of sciences , volume=

    Overcoming catastrophic forgetting in neural networks , author=. Proceedings of the national academy of sciences , volume=. 2017 , publisher=

    2017

  24. [24]

    International conference on machine learning , pages=

    Continual learning through synaptic intelligence , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  25. [25]

    arXiv preprint arXiv:1812.00420 , year=

    Efficient lifelong learning with a-gem , author=. arXiv preprint arXiv:1812.00420 , year=

    Pith/arXiv arXiv

  26. [26]

    Continual Learning with Tiny Episodic Memories , author=

  27. [27]

    arXiv preprint arXiv:1708.01547 , year=

    Lifelong learning with dynamically expandable networks , author=. arXiv preprint arXiv:1708.01547 , year=

    Pith/arXiv arXiv

  28. [28]

    International conference on machine learning , pages=

    Overcoming catastrophic forgetting with hard attention to the task , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  29. [29]

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

    Der: Dynamically expandable representation for class incremental learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  30. [30]

    arXiv preprint arXiv:2505.12335 , year=

    Is artificial intelligence generated image detection a solved problem? , author=. arXiv preprint arXiv:2505.12335 , year=

    arXiv

  31. [31]

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

    A continual deepfake detection benchmark: Dataset, methods, and essentials , author=. Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

  32. [32]

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

    D3QE: Learning Discrete Distribution Discrepancy-aware Quantization Error for Autoregressive-Generated Image Detection , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  33. [33]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=

    NeuroRenderedFake: A Challenging Benchmark to Detect Fake Images Generated by Advanced Neural Rendering Methods , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=

  34. [34]

    arXiv preprint arXiv:2509.03516 , year=

    Easier Painting Than Thinking: Can Text-to-Image Models Set the Stage, but Not Direct the Play? , author=. arXiv preprint arXiv:2509.03516 , year=

    arXiv

  35. [35]

    European conference on computer vision , pages=

    Microsoft coco: Common objects in context , author=. European conference on computer vision , pages=. 2014 , organization=

    2014

  36. [36]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

    2009

  37. [37]

    2024 , howpublished =

    2024

  38. [38]

    Advances in neural information processing systems , volume=

    Alias-free generative adversarial networks , author=. Advances in neural information processing systems , volume=

  39. [39]

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

    Blendface: Re-designing identity encoders for face-swapping , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  40. [40]

    European Conference on Computer Vision , pages=

    Infinite-id: Identity-preserved personalization via id-semantics decoupling paradigm , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  41. [41]

    Advances in Neural Information Processing Systems , volume=

    The gan is dead; long live the gan! a modern gan baseline , author=. Advances in Neural Information Processing Systems , volume=

  42. [42]

    2023 , publisher=

    Inswap , author=. 2023 , publisher=

    2023

  43. [43]

    2023 , howpublished =

    2023

  44. [44]

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

    Semantic image synthesis with spatially-adaptive normalization , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  45. [45]

    arXiv preprint arXiv:1809.11096 , year=

    Large scale GAN training for high fidelity natural image synthesis , author=. arXiv preprint arXiv:1809.11096 , year=

    Pith/arXiv arXiv

  46. [46]

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

    Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  47. [47]

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

    Text-to-3d using gaussian splatting , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  48. [48]

    2025 , howpublished =

    2025

  49. [49]

    2025 , howpublished =

    Sharon, David and Brichtova, Nicole , title =. 2025 , howpublished =

    2025

  50. [50]

    European conference on computer vision , pages=

    Gaussctrl: Multi-view consistent text-driven 3d gaussian splatting editing , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  51. [51]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  52. [52]

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

    PLAN: Proactive Low-Rank Allocation for Continual Learning , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  53. [53]

    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=

  54. [54]

    Advances in neural information processing systems , volume=

    Genimage: A million-scale benchmark for detecting ai-generated image , author=. Advances in neural information processing systems , volume=

  55. [55]

    Abril and Robert Plant

    Patricia S. Abril and Robert Plant. The patent holder's dilemma: Buy, sell, or troll?. Communications of the ACM. doi:10.1145/1188913.1188915

    arXiv

  56. [56]

    Deciding equivalances among conjunctive aggregate queries

    Sarah Cohen and Werner Nutt and Yehoshua Sagic. Deciding equivalances among conjunctive aggregate queries. doi:10.1145/1219092.1219093

    arXiv

  57. [57]

    Special issue: Digital Libraries. 1996

    1996

  58. [58]

    Understanding Policy-Based Networking

    David Kosiur. Understanding Policy-Based Networking

  59. [61]

    doi:10.1007/3-540-09237-4

    The title of book two. doi:10.1007/3-540-09237-4

    doi:10.1007/3-540-09237-4

  60. [62]

    Asad Z. Spector. Achieving application requirements. Distributed Systems. doi:10.1145/90417.90738

    arXiv

  61. [63]

    Douglass and David Harel and Mark B

    Bruce P. Douglass and David Harel and Mark B. Trakhtenbrot. Statecarts in use: structured analysis and object-orientation. Lectures on Embedded Systems. doi:10.1007/3-540-65193-4_29

    doi:10.1007/3-540-65193-4_29

  62. [64]

    Donald E. Knuth. The Art of Computer Programming, Vol. 1: Fundamental Algorithms (3rd. ed.)

  63. [65]

    Donald E. Knuth. The Art of Computer Programming

  64. [66]

    Structured Variational Inference Procedures and their Realizations (as incol)

    Dan Geiger and Christopher Meek. Structured Variational Inference Procedures and their Realizations (as incol). Proceedings of Tenth International Workshop on Artificial Intelligence and Statistics, The Barbados

  65. [67]

    Stan W. Smith. An experiment in bibliographic mark-up: Parsing metadata for XML export. Proceedings of the 3rd. annual workshop on Librarians and Computers

  66. [68]

    Catch me, if you can: Evading network signatures with web-based polymorphic worms

    Matthew Van Gundy and Davide Balzarotti and Giovanni Vigna. Catch me, if you can: Evading network signatures with web-based polymorphic worms. Proceedings of the first USENIX workshop on Offensive Technologies

  67. [69]

    Predicate Path expressions

    Sten Andler. Predicate Path expressions. Proceedings of the 6th. ACM SIGACT-SIGPLAN symposium on Principles of Programming Languages. doi:10.1145/567752.567774

    arXiv

  68. [70]

    LOGICS of Programs: AXIOMATICS and DESCRIPTIVE POWER

    David Harel. LOGICS of Programs: AXIOMATICS and DESCRIPTIVE POWER

  69. [71]

    Anisi , title =

    David A. Anisi , title =

  70. [72]

    Clarkson

    Kenneth L. Clarkson. Algorithms for Closest-Point Problems (Computational Geometry)

  71. [73]

    Introduction to Bayesian Statistics

    Harry Thornburg. Introduction to Bayesian Statistics. 2001

    2001

  72. [74]

    CLIFFORD: a Maple 11 Package for Clifford Algebra Computations, version 11

    Rafal Ablamowicz and Bertfried Fauser. CLIFFORD: a Maple 11 Package for Clifford Algebra Computations, version 11. 2007

    2007

  73. [75]

    Stats and Analysis

    Poker-Edge.Com. Stats and Analysis. 2006

    2006

  74. [76]

    A more perfect union

    Barack Obama. A more perfect union

  75. [77]

    The fountain of youth

    Joseph Scientist. The fountain of youth

  76. [78]

    Solder man

    Dave Novak. Solder man. ACM SIGGRAPH 2003 Video Review on Animation theater Program: Part I - Vol. 145 (July 27--27, 2003). doi:10.945/woot07-S422

    2003

  77. [79]

    Interview with Bill Kinder: January 13, 2005

    Newton Lee. Interview with Bill Kinder: January 13, 2005. Comput. Entertain. doi:10.1145/1057270.1057278

    arXiv 2005

  78. [80]

    The Enabling of Digital Libraries

    Bernard Rous. The Enabling of Digital Libraries. Digital Libraries

  79. [82]

    (new) Finding minimum congestion spanning trees , journal =

    Werneck, Renato and Setubal, Jo\. (new) Finding minimum congestion spanning trees , journal =. doi:10.1145/351827.384253 , acmid = 384253, publisher =

    arXiv

  80. [84]

    and Mei, Alessandro , title =

    Conti, Mauro and Di Pietro, Roberto and Mancini, Luigi V. and Mei, Alessandro , title =. Inf. Fusion , volume =. 2009 , issn =. doi:10.1016/j.inffus.2009.01.002 , acmid =

    doi:10.1016/j.inffus.2009.01.002 2009

Showing first 80 references.