Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate

T0 review · 3 major / 5 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 By replacing linear cross-attention updates with a nonlinear residual attention gate, CPE erases target concepts in text-to-image diffusion models while leaving remaining concepts essentially unchanged.

desk verdict Strong empirical concept erasing with a nonlinear gate; the theory is shakier than the results, and gate selectivity is asserted more than measured. read the letter →

arxiv 2506.22806 v1 pith:XTCK67QH submitted 2025-06-28 cs.CV cs.LG

Byung Hyun Lee , Sungjin Lim , Seunggyu Lee , Dong Un Kang , Se Young Chun This is my paper
classification cs.CVcs.LG
keywords concepterasingtext-to-imagediffusionresidualattentiongatecross-attentionfine-tuningadversarialrobustnessmachineunlearningforgettingpreventionStable
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that the standard recipe for concept erasing in text-to-image diffusion models—fine-tuning only the cross-attention layers—cannot in general preserve the wide range of concepts that are not being erased. The reason is linear: any weight update large enough to shift the target concept also raises the expected change in the cross-attention output for every remaining concept, measured over a Gaussian-mixture model of text embeddings. The proposed solution, Concept Pinpoint Eraser (CPE), replaces the linear weight update with a nonlinear Residual Attention Gate that reads the text embedding and suppresses the update for everything that is not the target, trained with an attention anchoring loss and an adversarial loop that hardens the gate against attack prompts. On benchmarks erasing 50 celebrities, 100 artistic styles, and explicit content, CPE deletes the targets while keeping remaining-concept metrics close to the unmodified model, outperforming prior methods on preservation.

What carries the argument

Residual Attention Gate (ResAG): a per-concept nonlinear module inserted into each cross-attention key/value projection, defined as $f_c(E) = A_c S(v_c^T E A_c)$ with $A_c = \sigma\left(\frac{(U_{1,c} E)^T (U_{2,c} E)}{\sqrt{m}}\right)$, where $S$ is a diagonal sigmoid, $U_{1,c}$ and $U_{2,c}$ are low-rank projections shared across all cross-attention layers, and $v_c$ is a concept vector. It also carries a low-rank weight update $\Delta W_c = U_{4,c}^T U_{3,c}$ added residually to the projection. The gate is what makes erasure selective: the attention factor $A_c$ detects the target concept in the embedding, the sigmoid diagonal $S$ reweights tokens, and the product $\Delta W_c E f_c(E)$ is active only when the gate opens, so remaining concepts see almost no change. The attention anchoring loss $L_{\text{att}} = \mathbb{E} \lVert R_{\text{tar}}(E_{\text{anc}}) \rVert_F$ minimizes the gate output on anchor embeddings, and RARE alternates between learning adversarial residual embeddings that reactivate the target and retraining the gate to block them.

What would settle it

Run UnlearnDiff or Ring-A-Bell on a CPE-erased model and probe with paraphrased target prompts such as 'the former president married to Hillary' when 'Bill Clinton' is erased: if any adversarial or paraphrased prompt makes the gate stay closed and the target concept reappears in the generated image, the selectivity claim is falsified. A quantitative version: measure the gate value $f_c(E)$ over a set of paraphrases of the target; a gate that opens for the target but stays at near zero for all non-target paraphrases supports the claim, while any high-confidence regeneration of a target from a held-out paraphrase contradicts it.

Watch

Extended reading notes

Core claim

CPE's central claim is that concept erasure should be a switch, not a direction. Fine-tuning key and value projections in cross-attention is equivalent to adding a linear map $\Delta W$ to the projections; Theorem 2 shows that under a Gaussian-mixture model of text embeddings the expected squared change for remaining concepts is $C_3 \lVert \Delta W \rVert_F^2 + \sum_{r=1}^{R} \pi_r \lVert \Delta W \mu_r \rVert_F^2$, so any linear update strong enough to erase the target inflates the disturbance of other concepts. CPE instead adds $f_c(E)$, a per-concept attention gate, so the effective update becomes $\Delta W_c E f_c(E)$; Corollary 1 shows that if the gate output $V_r$ is near zero for remaining concepts and large for the target, the expected disturbance vanishes. The gate is trained so that target embeddings are pushed toward a surrogate concept embedding while anchor embeddings are pinned by the attention anchoring loss, and the loop of learning adversarial residual embeddings and retraining the gate (RARE) extends the erasure to adversarial prompts. Across celebrities, artistic styles, and explicit content, the method reports near-complete deletion of targets with remaining-concept metrics effectively matching the original model.

Load-bearing premise

The whole preservation advantage rests on the gate reliably telling the target concept apart from look-alike non-targets across paraphrases and adversarial prompts, and the paper concedes that an ideal classifier on embeddings is infeasible, leaving the gate trained on a limited set of target and anchor embeddings.

Editorial extensions

If this is right

  • Erasing 50 celebrities at once leaves 100 non-target celebrities, 100 artistic styles, 64 characters, and COCO-30K nearly unchanged, where baselines degrade at least one domain.
  • The same setup erases 100 artistic styles with remaining domains effectively matching the original Stable Diffusion v1.4.
  • Explicit-content erasure lowers detected nudity on I2P to 40 total detections versus 111 for the best prior method, with the best CLIP score among methods on COCO-30K.
  • Against Ring-A-Bell and UnlearnDiff attacks on celebrities, I2P nudity, and Van Gogh style, CPE reports zero or near-zero attack success rates on several tasks.
  • The ResAG adds less than 0.01% of model parameters per concept and can train in about two minutes per concept without the adversarial loop.

Reading between the lines

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

  • The same gate mechanism could be applied to other attention-based generative models, such as video or audio diffusion, wherever text embeddings drive cross-attention, since the gate only needs the embedding and the projection outputs.
  • Because the gate is a function of the text embedding, it could double as a cheap detector: the gate value for a target concept gives a training-free prompt audit for whether an erased concept still influences generation.
  • The Gaussian-mixture argument suggests the preservation problem is inherent to any linear fine-tuning scheme, so closed-form methods like UCE and MACE may need a nonlinear component rather than better anchors to fix preservation.
  • The adversarial loop makes erasure a moving target; a natural extension is to certify the gate over a neighborhood of embeddings by bounding the gate's Lipschitz constant on the target boundary.
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

3 major / 5 minor

Summary. The paper studies concept erasing in text-to-image diffusion models. It first argues, via Theorems 1 and 2, that fine-tuning only the linear cross-attention (CA) projections cannot in general preserve diverse remaining concepts, because the expected change in the CA output for remaining concepts grows with the norm of the weight update. It then proposes Concept Pinpoint Eraser (CPE), which augments CA key/value projections with nonlinear Residual Attention Gates (ResAGs), trained with an erasing loss, an attention anchoring loss, and an adversarial-training scheme (RARE) over learnable residual embeddings. The experimental section compares CPE with FMN, ESD, UCE, MACE, RECE, and AdvUnlearn on erasing 50 celebrities, 100 artistic styles, and explicit content, reporting target-erasure and preservation metrics on celebrities, styles, characters, and COCO-30K, plus robustness to Ring-A-Bell and UnlearnDiff attacks.

Significance. If the reported results hold, CPE is a practical and parameter-efficient method: it adds less than 0.01% of model parameters per concept, and it shows unusually strong preservation of remaining concepts across several domains while remaining competitive in erasure strength. The evaluation is grounded in external detectors (GCD, NudeNet) and standard image-quality metrics (CLIP score, FID, KID), and the paper includes extensive ablations on ranks, loss weights, anchor selection, and the proposed components. The release of code is a further strength. The principal weakness is that the paper's central mechanism, the selectivity of the learned attention gate, is never directly measured; the theoretical motivation also contains a factor-of-m inconsistency that should be corrected.

major comments (3)
  1. [Section 3.1, Eq. (3); Appendix B, Theorem 2 proof] Theorem 2 is stated inconsistently with its proof. In the main text, Eq. (3) defines C3 = sum_{r=1}^R pi_r sigma_r^2, but the proof in Appendix B concludes C3 = m * sum_{r=1}^R pi_r sigma_r^2, where m is the token sequence length. The qualitative conclusion (growing ||Delta W||_F to erase targets also raises the expected change for remaining concepts) is unaffected by the factor m, but Theorem 2 is the paper's central theoretical motivation and the statement must agree with the proof. Please correct the theorem statement or the proof, and check whether the same issue affects the presentation of Corollary 1.
  2. [Section 3.2, Eq. (6); Section 3.3, Algorithm 1] The paper never directly measures the selectivity of the learned gate f_c(E), although this selectivity is the load-bearing mechanism behind the preservation claims. For example, the target list in Table D.1 contains 'Bill Clinton' and the remaining-concepts list in Table D.2 contains 'Bill Murray'; the paper argues that such similar pairs are exactly where the gate must discriminate. However, Algorithm 1 supervises only projection outputs, and no experiment reports gate activations on target vs. remaining concepts, on unseen paraphrase templates, or on the learned adversarial embeddings. Please add a direct gate analysis, e.g., histograms or mean diagonal gate values for target prompts, similar non-target prompts, and paraphrases, and report worst-case per-concept preservation failures. Without this evidence, the claimed advantage of the nonlinear gate over a linear CA edit is only indirectly supported.
  3. [Appendix D.3, Tables D.3 and D.6] The anchor concepts used in the attention anchoring loss are selected from an LLM-generated pool (500 celebrities and 1,734 artistic styles), and the remaining-concept evaluation sets comprise 100 celebrities and 100 artistic styles. The manuscript does not state whether the selected anchor concepts overlap with the evaluation lists. Because Eq. (8) directly minimizes the CA-output change on anchor concepts, any overlap between anchors and the evaluation set would inflate the reported preservation numbers. Please report the overlap, and if nonempty, re-report preservation metrics on a held-out subset with all anchor concepts excluded.
minor comments (5)
  1. [Section 3.1, Eq. (2)] The constant C_1^h is typed as proportional to sqrt(m) - s_1, but the Appendix B proof uses sqrt(m) - 1. Since s_1 is only introduced later as the rank of the attention gate, the main-text equation appears to contain a typo and should be corrected.
  2. [Table 2] The table header reads 'atistic styles erasure' and should be 'artistic styles erasure'.
  3. [Sections 4.1 and 4.2] The figure references appear to be swapped: Section 4.1 refers to 'Figure 5' when discussing celebrities erasure, which should be Figure 3, and Section 4.2 refers to 'Figure 3' when discussing artistic styles erasure, which should be Figure 4.
  4. [Appendix B, Theorem 2 proof] The phrase 'simliar' should be 'similar', and the displayed equation after Eq. (B.4) has the mixture weight pi_r placed outside the sum over r, which makes the derivation harder to follow; please correct the typesetting.
  5. [Figure 1 caption] The caption contains 'adatively transmit' and should read 'adaptively transmit'.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: preservation numbers on same-domain remaining concepts include concepts explicitly trained as attention anchors; erasing, robustness, and held-out domains remain independent.

  1. fitted input called prediction [Section 3.3 Eq. (8); Section 4.1; Appendix D.3.1; Tables D.2/D.5 vs D.10]
    "To prevent undesirable degradation on remaining concepts by erasing loss, we propose an attention anchoring loss ... L_att(E_anc) = E_{E_anc} ||R_tar(E_anc)||_F. For remaining concepts, we considered three domains: 100 celebrities and 100 artistic styles from (Lu et al., 2024) ... we select 50 anchor concepts with high cosine similarity to the selected target celebrity in text embeddings, from the anchor concept pool."

    The anchor concepts optimized by Eq. (8) are drawn from the same pools later reported as 'remaining concepts' in Tables 1-2. For target Hugh Jackman, Table D.10 lists anchors Keanu Reeves, Ben Affleck, Jon Hamm, and Jason Statham, all of which appear in Table D.2's remaining 100 celebrities. For target Claude Monet, anchors include Camille Pissarro, Paul Cezanne, Henri Matisse, and Albert Bierstadt, all in Table D.5's remaining 100 styles. Since Eq. (8) directly minimizes the gate output on these exact embeddings, the reported KID/CS/ACC preservation on those concepts measures the training objective rather than an independent prediction of broad-concept retention.

full rationale

Most of the derivation chain is self-contained. Theorems 1-2 and Corollary 1 are conditional mathematical statements under an explicitly stated Gaussian-mixture assumption; they do not presuppose the method's outputs. The gate design is presented as an ansatz ('we are inspired by the mechanism of attention gates'), and Proposition 1's ideal classifier is explicitly admitted infeasible, so no uniqueness or external-support claim is smuggled in. The self-citations (Lee et al. 2023, 2024; Lim et al. 2022) appear only in related-work or augmentation contexts and are not load-bearing. Target erasure and attack robustness are measured with external detectors (GCD, NudeNet, Ring-A-Bell, UnlearnDiff) against standard baselines, and the held-out preservation domains (COCO-30K/COCO-1K, 64 characters, cross-domain remaining sets) are independent. The one concrete circularity is an evaluation overlap: the anchor concepts trained by the attention anchoring loss overlap with the reported same-domain remaining-concept test sets. Because Eq. (8) literally minimizes the gate response on those anchors, reporting them as evidence of preservation partially re-states the training objective. This is a non-fatal, partial circularity: the central erasing claim and the held-out preservation results retain independent content, so a moderate score of 4 is appropriate.

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

The method's success depends on several hand-chosen hyperparameters (η, λ, ranks, number of adversarial stages) and domain assumptions (Gaussian-mixture text embeddings; gate approximates an ideal classifier; anchor concepts represent remaining distribution). None of these are derived from first principles, but they are standard for empirical deep learning papers and are ablated.

free parameters (6)
  • η (erasing loss intensity) = 0.3 (celebrities), 0.5 (artistic styles), 3.0 (explicit content)
    Hyperparameter controlling how aggressively target embeddings are pushed toward surrogate; chosen by validation, affects erasure/preservation trade-off (Appendix E).
  • λ (attention anchoring loss weight) = 1e5 (celebrities), 1e4 (artistic styles/explicit)
    Controls preservation of anchor concepts; ablation shows trade-off with erasure (Table E.3).
  • s1 (attention gate rank) = 16, 16, 64
    Rank of attention matrices U1,U2; chosen per task, ablation shows low rank suffices (Table 5).
  • s2 (low-rank residual rank) = 1, 1, 4
    Rank of U3,U4; ablation shows rank 1 works for celebrities (Table E.1).
  • N (number of adversarial embeddings) = 16, 16, 64
    Number of learned residual embeddings for RARE; increases with task difficulty (Appendix D.3).
  • S (number of adversarial stages) = 5, 10, 20
    Iterations of erasing/adversarial training; chosen per task (Appendix D.3).
assumptions (4)
  • domain assumption Text embeddings of remaining concepts follow a Gaussian mixture model
    Used in Theorem 2 to derive the expectation of ||ΔW E_rem||_F^2 and the 'dilemma' that increasing ||ΔW||_F increases the bound. Actual CLIP embeddings likely not Gaussian; this is a simplification for analysis (Section 3.1).
  • domain assumption Change in cross-attention output propagates to diffusion model output
    Assumed at start of Section 3.1: 'we assume that the change in a CA layer output can induce the change in the diffusion model output from the original output.' This bridges the bound to forgetting but is not proven.
  • ad hoc to paper The learned attention gate can approximate the ideal concept classifier
    Proposition 1 defines an ideal α(E) that is infeasible; the paper assumes a learned attention gate can approximate this behavior well enough to separate target from similar non-target concepts (Section 3.2).
  • standard math Lipschitz constants and norms M1, M2 are finite and computable
    Used in Theorem 1 proof; standard analytical assumptions for bounding softmax attention outputs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate." pith.science (2026). https://pith.science/paper/XTCK67QH

@misc{pith2026250622806,
  author       = {Pith},
  title        = {Pith review of: Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XTCK67QH}},
  note         = {Machine review of arXiv:2506.22806}
}
read the original abstract

Remarkable progress in text-to-image diffusion models has brought a major concern about potentially generating images on inappropriate or trademarked concepts. Concept erasing has been investigated with the goals of deleting target concepts in diffusion models while preserving other concepts with minimal distortion. To achieve these goals, recent concept erasing methods usually fine-tune the cross-attention layers of diffusion models. In this work, we first show that merely updating the cross-attention layers in diffusion models, which is mathematically equivalent to adding \emph{linear} modules to weights, may not be able to preserve diverse remaining concepts. Then, we propose a novel framework, dubbed Concept Pinpoint Eraser (CPE), by adding \emph{nonlinear} Residual Attention Gates (ResAGs) that selectively erase (or cut) target concepts while safeguarding remaining concepts from broad distributions by employing an attention anchoring loss to prevent the forgetting. Moreover, we adversarially train CPE with ResAG and learnable text embeddings in an iterative manner to maximize erasing performance and enhance robustness against adversarial attacks. Extensive experiments on the erasure of celebrities, artistic styles, and explicit contents demonstrated that the proposed CPE outperforms prior arts by keeping diverse remaining concepts while deleting the target concepts with robustness against attack prompts. Code is available at https://github.com/Hyun1A/CPE

Figures

Figures reproduced from arXiv: 2506.22806 by the authors.

Figure 1
Figure 1. (a) Comparison of fine-tuning approaches for concept erasing. Previous methods could [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) Architecture of ResAG module in CA layers for selectively erasing a target concept [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results of our CPE and baselines on multiple concepts erasing. We erased 50 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative results on artistic styles erasure. We erased 100 artistic styles at once. It shows [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results of CPE and baselines on robustness to adversarial attacks by Unlearn [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoders

    cs.CV 2025-09 conditional novelty 6.0 of 10

    A supervised sparse autoencoder binds each concept to a single neuron, letting Stable Diffusion erase a concept by steering one latent.

Reference graph

Works this paper leans on

79 extracted references · 60 canonical work pages · cited by 1 Pith paper

  1. [1]

    Persistent anti-muslim bias in large language models

    Abubakar Abid, Maheen Farooqi, and James Zou. Persistent anti-muslim bias in large language models. AAAI/ACM on AI, Ethics, and Society, 2021

    work page 2021

  2. [2]

    Bottom-up and top-down attention for image captioning and visual question answering

    Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. CVPR, 2018

    work page 2018

  3. [3]

    Neural machine translation by jointly learning to align and translate

    Dzmitry Bahdanau, Kyung Hyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. ICLR, 2015

    work page 2015

  4. [4]

    Nudenet: Neural nets for nudity detection and censoring

    Praneeth Bedapudi. Nudenet: Neural nets for nudity detection and censoring. https://nudenet.notai.tech/, 2022

    work page 2022

  5. [5]

    Large image datasets: A pyrrhic win for computer vision? WACV, 2021

    Abeba Birhane and Vinay Uday Prabhu. Large image datasets: A pyrrhic win for computer vision? WACV, 2021

    work page 2021

  6. [6]

    Muse: Text-to-image generation via masked generative transformers

    Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. ICML, 2023

    work page 2023

  7. [7]

    Crossvit: Cross-attention multi-scale vision transformer for image classification

    Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. ICCV, 2021

    work page 2021

  8. [8]

    Word2vec

    Kenneth Ward Church. Word2vec. Natural Language Engineering, 2017

    work page 2017

Show all 79 references
  1. [9]

    Regavae: A retrieval-augmented gaussian mixture variational auto-encoder for language modeling

    Jingcheng Deng, Liang Pang, Huawei Shen, and Xueqi Cheng. Regavae: A retrieval-augmented gaussian mixture variational auto-encoder for language modeling. EMNLP, 2023

    2023

  2. [10]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021

    2021

  3. [11]

    Measuring and mitigating unintended bias in text classification

    Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. Measuring and mitigating unintended bias in text classification. AAAI/ACM on AI, Ethics, and Society, 2018

    2018

  4. [12]

    Compositional visual generation with energy based models

    Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. NeurIPS, 2020

    2020

  5. [13]

    Unsupervised learning of compositional energy concepts

    Yilun Du, Shuang Li, Yash Sharma, Josh Tenenbaum, and Igor Mordatch. Unsupervised learning of compositional energy concepts. Neurips, 2021

    2021

  6. [14]

    Tweedie’s formula and selection bias

    Bradley Efron. Tweedie’s formula and selection bias. Journal of the American Statistical Association, 106 0 (496): 0 1602--1614, 2011

    2011

  7. [15]

    Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation

    Chongyu Fan, Jiancheng Liu, Yihua Zhang, Dennis Wei, Eric Wong, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. ICLR, 2024

    2024

  8. [16]

    Erasing concepts from diffusion models

    Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. ICCV, 2023

    2023

  9. [17]

    Unified concept editing in diffusion models

    Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzy \'n ska, and David Bau. Unified concept editing in diffusion models. WACV, 2024

    2024

  10. [18]

    Reliable and efficient concept erasure of text-to-image diffusion models

    Chao Gong, Kai Chen, Zhipeng Wei, Jingjing Chen, and Yu-Gang Jiang. Reliable and efficient concept erasure of text-to-image diffusion models. arXiv preprint arXiv:2407.12383, 2024

    2024 arXiv

  11. [19]

    Giphy celebrity detector

    Nick Hasty, Ihor Kroosh, Dmitry Voitekh, and Dmytro Korduban. Giphy celebrity detector. https://github.com/Giphy/celeb-detection-oss, 2024

    2024

  12. [20]

    Selective amnesia: A continual learning approach to forgetting in deep generative models

    Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. NeurIPS, 2023

    2023

  13. [21]

    Clipscore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. EMNLP, 2021

    2021

  14. [22]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017

    2017

  15. [23]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020

    2020

  16. [24]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 2021

    2021

  17. [25]

    Receler: Reliable concept erasing of text-to-image diffusion models via lightweight erasers

    Chi-Pin Huang, Kai-Po Chang, Chung-Ting Tsai, Yung-Hsuan Lai, and Yu-Chiang Frank Wang. Receler: Reliable concept erasing of text-to-image diffusion models via lightweight erasers. arXiv preprint arXiv:2311.17717, 2023

    2023 arXiv

  18. [26]

    Robo-writers: the rise and risks of language-generating ai

    Matthew Hutson. Robo-writers: the rise and risks of language-generating ai. Nature, 591 0 (7848): 0 22--25, 2021

    2021

  19. [27]

    Image synthesis style studies

    Surea I, Proxima Centauri B, Erratica, and Stephen Young. Image synthesis style studies. https://www.aiartapps.com/ai-art-apps/image-synthesis-style-studies

  20. [28]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015

    2015

  21. [29]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. ICCV, 2023

    2023

  22. [30]

    Overcoming catastrophic forgetting in neural networks

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of scien...

    2017

  23. [31]

    Ablating concepts in text-to-image diffusion models

    Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. ICCV, 2023

    2023

  24. [32]

    Nsfw detection machine learning model

    Gant Laborde. Nsfw detection machine learning model. https://github.com/GantMan/nsfw_model, 2020

    2020

  25. [33]

    Online continual learning on hierarchical label expansion

    Byung Hyun Lee, Okchul Jung, Jonghyun Choi, and Se Young Chun. Online continual learning on hierarchical label expansion. ICCV, 2023

    2023

  26. [34]

    Doubly perturbed task free continual learning

    Byung Hyun Lee, Min-hwan Oh, and Se Young Chun. Doubly perturbed task free continual learning. AAAI, 2024

    2024

  27. [35]

    Self-attention enhanced selective gate with entity-aware embedding for distantly supervised relation extraction

    Yang Li, Guodong Long, Tao Shen, Tianyi Zhou, Lina Yao, Huan Huo, and Jing Jiang. Self-attention enhanced selective gate with entity-aware embedding for distantly supervised relation extraction. AAAI, 2020

    2020

  28. [36]

    Noisy feature mixup

    Soon Hoe Lim, N Benjamin Erichson, Francisco Utrera, Winnie Xu, and Michael W Mahoney. Noisy feature mixup. ICLR, 2022

    2022

  29. [37]

    Gpt understands, too

    Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. AI Open, 2023

    2023

  30. [38]

    Sgdr: Stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. ICLR, 2016

    2016

  31. [39]

    Tf-icon: Diffusion-based training-free cross-domain image composition

    Shilin Lu, Yanzhu Liu, and Adams Wai-Kin Kong. Tf-icon: Diffusion-based training-free cross-domain image composition. ICCV, 2023

    2023

  32. [40]

    Mace: Mass concept erasure in diffusion models

    Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. CVPR, 2024

    2024

  33. [41]

    One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications

    Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications. CVPR, 2024

    2024

  34. [42]

    A. M. Mathai and Serge B. Provost. Quadratic Forms in Random Variables: Theory and Applications, volume 126 of Statistics: Textbooks and Monographs. Marcel Dekker, New York, 1992. ISBN 9780824786915

    1992

  35. [43]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models

    Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. ICML, 2022

    2022

  36. [44]

    Stable diffusion 1 vs 2 - what you need to know

    Ryan O'connor. Stable diffusion 1 vs 2 - what you need to know. https://www.assemblyai.com/blog/stable-diffusion-1-vs-2-what-you-need-to-know/, 2022

    2022

  37. [45]

    Attention u-net: Learning where to look for the pancreas

    Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. MIDL, 2022

    2022

  38. [46]

    Editing implicit assumptions in text-to-image diffusion models

    Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. ICCV, 2023

    2023

  39. [47]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. ICML, 2021

    2021

  40. [48]

    Hierarchical text-conditional image generation with clip latents

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

    2022 arXiv

  41. [49]

    Red-teaming the stable diffusion safety filter

    Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramer. Red-teaming the stable diffusion safety filter. NeurIPS ML Safety Workshop, 2022

    2022

  42. [50]

    Stable diffusion 2.0 release

    Robin Rombach. Stable diffusion 2.0 release. https://stability.ai/news/stable-diffusion-v2-release, 2022

    2022

  43. [51]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2022 a

    2022

  44. [52]

    Stable diffusion v1 model card

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. Stable diffusion v1 model card. https://huggingface.co/CompVis/stable-diffusion-v1-4, 2022 b

    2022

  45. [53]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. MICCAI, 2015

    2015

  46. [54]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022

    2022

  47. [55]

    Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models

    Patrick Schramowski, Manuel Brack, Bj \"o rn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. CVPR, 2023

    2023

  48. [56]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs

    Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021

    2021 arXiv

  49. [57]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022

    2022

  50. [58]

    Active learning for convolutional neural networks: A core-set approach

    Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. ICLR, 2018

    2018

  51. [59]

    Diffusion art or digital forgery? investigating data replication in diffusion models

    Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. CVPR, 2023

    2023

  52. [60]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. NeurIPS, 2020

    2020

  53. [61]

    Demystifying mmd gans

    JD Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. ICLR, 2018

    2018

  54. [62]

    Language identification using gaussian mixture model tokenization

    Pedro A Torres-Carrasquillo, Douglas A Reynolds, and John R Deller. Language identification using gaussian mixture model tokenization. ICASSP, 2002 a

    2002

  55. [63]

    Approaches to language identification using gaussian mixture models and shifted delta cepstral features

    Pedro A Torres-Carrasquillo, Elliot Singer, Mary A Kohler, Richard J Greene, Douglas A Reynolds, and John R Deller Jr. Approaches to language identification using gaussian mixture models and shifted delta cepstral features. Interspeech, 2002 b

    2002

  56. [64]

    Ring-a-bell! how reliable are concept removal methods for diffusion models? ICLR, 2024

    Yu-Lin Tsai, Chia-yi Hsu, Chulin Xie, Chih-hsun Lin, Jia You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? ICLR, 2024

    2024

  57. [65]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017

    2017

  58. [66]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017

    2017

  59. [67]

    Deep gaussian mixture models

    Cinzia Viroli and Geoffrey J McLachlan. Deep gaussian mixture models. Statistics and Computing, 29: 0 43--51, 2019

    2019

  60. [68]

    Universality and limitations of prompt tuning

    Yihan Wang, Jatin Chauhan, Wei Wang, and Cho-Jui Hsieh. Universality and limitations of prompt tuning. NeurIPS, 2024

    2024

  61. [69]

    Visual-language prompt tuning with knowledge-guided context optimization

    Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. CVPR, 2023

    2023

  62. [70]

    Forget-me-not: Learning to forget in text-to-image diffusion models

    Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023 a

    2023 arXiv

  63. [71]

    mixup: Beyond empirical risk minimization

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. ICLR, 2017

    2017

  64. [72]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. ICCV, 2023 b

    2023

  65. [73]

    To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images

    Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868, 2023 c

    arXiv

  66. [74]

    Defensive unlearning with adversarial training for robust concept erasure in diffusion models

    Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234, 2024

    2024 arXiv

  67. [75]

    Learning to prompt for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022

    2022

  68. [76]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  69. [77]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  70. [78]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  71. [79]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

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