Pith. sign in

REVIEW 4 major objections 6 minor 66 references

Logo unlearning fails to erase brand marks without wrecking images

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 23:03 UTC pith:Z6DMZEBO

load-bearing objection A solid benchmark with a real finding, but the local metrics need a detector-evasion fix before the numbers are trustworthy. the 4 major comments →

arxiv 2607.24101 v2 pith:Z6DMZEBO submitted 2026-07-27 cs.CV cs.AI

LU-500: A Logo Benchmark for Concept Unlearning

Keyu Li , Jin Gao , Jialing Zhang , Dequan Wang This is my paper
classification cs.CV cs.AI
keywords logo unlearningconcept unlearningtext-to-image modelsbenchmarklocalized concept erasureFortune Global 500Stable Diffusionprompt rewriting
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.

This paper introduces LU-500, a benchmark of nearly 10,000 text-image pairs built from Fortune Global 500 companies, to test whether concept-unlearning methods can remove a small, localized logo while preserving the rest of a generated scene. Across explicit prompts that name the logo and implicit prompts that evoke it through context, the evaluated inference-time methods (NP, SLD, SEGA) and two fine-tuning methods (ESD, Forget-Me-Not) all show the same pattern: stronger intervention blurs or deforms the logo but also degrades the surrounding image, while gentler intervention leaves a recognizable brand trace. The paper proposes a multi-grained metric protocol that separates local erasure from global preservation, and introduces ProLU, a prompt-rewriting baseline that achieves strong local removal but at the cost of changing the scene. The authors argue that logo unlearning needs spatially aware controls, such as SSIM-guided constraints, rather than purely global concept suppression.

Core claim

The central claim is that logos constitute a distinct and harder failure mode for concept unlearning: a tiny, precisely drawn mark can carry the entire protected concept, and current methods cannot remove it without altering non-target content. On LU-500, CLIPScore (text-logo alignment) remains well above the ideal zero for all evaluated methods, and increasing safety guidance in SLD or SEGA improves local metrics like LogoSSIM only while ImageScore and ImageSSIM plummet. ProLU, a prompt-space multi-agent baseline, achieves the strongest local erasure (CLIPScore 0.26) but drops global fidelity, demonstrating that semantic rewriting is not a substitute for weight-level disentanglement. Correl

What carries the argument

The evaluation protocol is built on two components: OWLv2, an open-vocabulary detector that localizes logo regions in generated images, and five metrics computed over those regions and the full image—CLIPScore (text-to-logo semantic alignment), LogoScore and LogoSSIM (local logo change in latent and pixel space), and ImageScore and ImageSSIM (global scene preservation). ProLU, the diagnostic baseline, orchestrates three LLM agents—Remover, Reflector, and Checker—to iteratively rewrite prompts and strip logo-inducing semantics while trying to keep the scene intact.

Load-bearing premise

The conclusion that methods struggle to remove logos rests on treating a logo that the OWLv2 detector no longer finds as fully erased; if an intervention merely deforms or hides the mark so the detector misses it, the local metrics record success even though the brand is still visually present.

What would settle it

Take a sample of post-unlearning images that the protocol scores as zero residual logo and run human raters (or a second, independently trained detector) to check whether the brand mark is still recognizable. The paper reports 98% detection accuracy only on pre-unlearning images (§3.3.1), so this check targets the unmeasured confound of detector evasion.

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

If this is right

  • Binary success rates are insufficient for logo unlearning; the local-global trade-off is only visible when erasure and preservation are measured separately.
  • Inference-time guidance methods do not inherently localize to small brand marks, so improving them will require spatial awareness, not just stronger safety scaling.
  • Prompt rewriting can remove many explicit and contextual logo triggers, but it changes the input description and cannot guarantee weight-level forgetting.
  • Larger and high-complexity logo regions resist structural alteration, suggesting that interventions should be guided by region-specific constraints such as SSIM.
  • Newer text-to-image models render logos with higher fidelity, making unlearning harder and benchmarks like LU-500 more necessary.

Where Pith is reading between the lines

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

  • The protocol assigns zero residual logo whenever OWLv2 detects no box after unlearning; if an intervention deforms or hides the mark so the detector misses it, local metrics would record success even though a human could still recognize the brand—this detector-evasion confound should be tested on post-intervention images.
  • The explicit/implicit track split could be extended to adversarial prompts, multilingual brand variants, or long-tail regional logos, which the paper itself lists as future work but does not evaluate.
  • The paper's ranking of the most entangled logos (e.g., Hyundai Motor, China Aerospace Science & Industry, Panasonic) provides a natural stress-test subset: future unlearning methods could be required to pass those cases first.
  • If SSIM-guided spatial constraints prove effective, they could be implemented as attention-mask losses in fine-tuning methods, directly protecting background regions during weight updates.

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

4 major / 6 minor

Summary. The paper introduces LU-500, a benchmark for logo unlearning in text-to-image models, built from Fortune Global 500 companies and containing nearly 10,000 human-validated prompt/logo pairs across an explicit track (LUex-500) and an implicit contextual track (LUim-500). It proposes a multi-grained evaluation protocol with five metrics: local logo erasure (CLIPScore, LogoScore, LogoSSIM) and global image preservation (ImageScore, ImageSSIM). The authors evaluate inference-time methods (NP, SLD, SEGA) and fine-tuning methods (ESD, Forget-Me-Not), and introduce ProLU, a prompt-space multi-agent diagnostic baseline. The central claim is that current unlearning methods struggle to remove localized logo evidence without changing non-target content, and that prompt-space rewriting is not a substitute for weight-level disentanglement.

Significance. If the claims are robust, LU-500 fills a genuine gap: existing unlearning benchmarks target whole-image concepts, while logos require localized, semantically entangled removal. The release of a large, human-validated dataset and the multi-grained metric protocol are potentially valuable community resources. The paper also explicitly scopes its conclusions to the evaluated methods and backbones, which is appropriate. However, the central empirical claim rests on detector-gated local metrics that can conflate detector evasion with true erasure, and the quantitative tables lack variance information. These issues are addressable but currently compromise the strength of the main conclusion.

major comments (4)
  1. [§3.3.2] The zero-assignment rule for undetected logos is a load-bearing confound. The text states: 'When no logo is detected after unlearning, local metrics are assigned zero, corresponding to no measurable residual logo region under the detector.' The 98% OWLv2 accuracy was verified on pre-unlearning generated images (§3.3.1), not on post-intervention images. Since SLD/SEGA distort, deform, or hide logos, the detector may miss a still-present brand mark, and CLIPScore, LogoScore, and LogoSSIM collapse to zero, recording 'perfect' erasure. This directly threatens the claim that 'the evaluated methods struggle to remove logo evidence.' The paper should report the no-detection rate separately, validate OWLv2 on post-unlearning images (e.g., with human verification or a second detector), and either conservatively impute non-zero local scores when the detector misses or discuss the evasion issue exp
  2. [Table 1 / §4.1] All quantitative results are single point estimates with no error bars, confidence intervals, or variance across seeds or companies, and the text says the random seed is fixed. For a benchmark claiming to characterize 'systemic inefficacy' across ~10,000 prompts and 500 companies, this is insufficient: differences between methods and across SLD/SEGA variants may not be statistically significant. The paper should provide per-company or per-prompt variance, bootstrap confidence intervals, or multiple-seed results, and ideally report detection-rate statistics as well. This is necessary to support the quantitative comparison in Table 1.
  3. [Table 4 / §4.5] The 'Relative Contribution' column is internally inconsistent with the reported CLIPScore reductions. From the table: ProLU gives 16.76% reduction; without Remover, reduction is 3.07%; without Reflector, reduction is 15.15%; without Checker, reduction is 16.76%. Attributing the Remover's contribution as (16.76−3.07)/16.76 ≈ 81.7% and the Reflector's as (16.76−15.15)/16.76 ≈ 9.6%, not 83.2% and 16.8% as printed. The Checker contribution is 0%. This appears to be an arithmetic error in a table used to characterize the ProLU framework; it should be corrected or the computation method clarified.
  4. [§4.3 / Table 3] The fine-tuning comparison places ESD on SD-v1-4 and Forget-Me-Not on SD-v2-1-base, i.e., different backbones, and ProLU is evaluated on each backbone separately. The table caption and text compare ProLU against both methods ('ProLU outperforms ESD and Forget-Me-Not'), but such cross-backbone rankings are not apples-to-apples because the base generative model differs. The paper explicitly scopes conclusions, which is good, but the bold/underline rankings in Table 3 still imply a fair comparison. Please present results separately per backbone, or add matched-backbone baselines, and avoid cross-method ranking unless the backbone is controlled. This also affects the abstract's mention of 'compatible fine-tuning-based methods such as ESD and Forget-Me-Not.'
minor comments (6)
  1. [Figure 8 vs §3.1] The text in §3.1 says prompts were generated with GPT-5, while the Figure 8 caption says GPT-4o. Please reconcile this inconsistency.
  2. [Table 1] The caption mentions 'Original CLIP-Text scores 0.32 on explicit (Ex) and implicit (Im) tracks,' but the table does not include a 'Before' row. Showing the pre-unlearning local scores would help readers interpret the reductions.
  3. [Figure 6 / §4.4] Figure 6 is captioned 'SEGA unlearning efficacy,' but §4.4 discusses correlations with unlearning behavior generally. Clarify that the correlations are computed for SEGA only (or for all methods) and add confidence intervals or significance tests for the Pearson values.
  4. [§3.3.1] The '98% detection accuracy in this verification' is stated without sample size, verification protocol, or confidence interval. Given that this detector is the gate for all local metrics, this validation should be described fully.
  5. [Algorithm 1] Algorithm 1's metric naming and notation are hard to follow (e.g., M_E1 is used for both pre- and post-unlearning text-logo alignment; the equation formatting is broken). Please rewrite it more cleanly.
  6. [References and typos] There are typographical issues (e.g., 'V olvo' in Figures 10–11, 'dalle3' in the references) and the section title 'Concept Unlearn' in §2.2 is ungrammatical. Reference [55] (Lin divergence) does not appear to be the standard citation for fractal dimension; please check.

Circularity Check

0 steps flagged

No circularity found: LU-500 is a measurement benchmark with externally defined prompts, metrics, and baselines.

full rationale

LU-500's central claim is that evaluated concept-unlearning methods struggle to remove logo evidence without degrading non-target content. This claim rests on measured comparisons between pre- and post-intervention images using defined metrics (CLIPScore, LogoScore, LogoSSIM, ImageScore, ImageSSIM) and an external detector (OWLv2). The benchmark construction is a human-AI prompt pipeline with filtering; no parameter is fitted to the target result, and no equation defines the outcome in terms of itself. The metrics are computed from actual image pairs, and the local-global trade-off is an empirical observation, not a definitional consequence. ProLU is introduced explicitly as a diagnostic prompt-space baseline and is not used to define benchmark success; its performance is measured with the same external metrics. The self-citations to prior multi-agent work motivate the design of ProLU but are not load-bearing for the benchmark's conclusions. The concern that local metrics assign zero when the detector misses the logo after unlearning is a measurement-validity issue, not circularity: it does not make the reported scores equivalent to the inputs by construction. No fitted parameter is rebranded as a prediction, no uniqueness theorem is imported from the authors' prior work, and no known result is renamed as a new contribution. The paper's stated limitations further scope its claims without circularly relying on its own conclusions. Therefore no significant circularity is present.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

LU-500's central claims are empirical, so the ledger records measurement and tooling assumptions rather than fitted parameters. No parameters were fit to the benchmark conclusions; the SLD/SEGA guidance strengths are explicit sweeps, and the unstated OWLv2 confidence threshold is a hidden degree of freedom. The main risk is that the detector, CLIP, and SSIM are treated as faithful proxies for logo presence and scene preservation.

free parameters (2)
  • OWLv2 detection confidence threshold = not reported
    Hidden degree of freedom: the threshold determines which boxes are treated as logos and therefore affects every local metric; no sensitivity analysis is given.
  • SLD/SEGA guidance strengths = see Table 2; e.g., SLD_v1 safety_guidance=2000, threshold=0.025; SEGA_v1 threshold=0.99
    Hand-selected intervention intensities used to sweep the local/global trade-off; conclusions are qualitative but exact scores depend on these choices.
axioms (5)
  • domain assumption OWLv2 with company-name text query reliably localizes logo regions in generated images.
    Used as evaluator oracle (§3.3.1); manual verification only on pre-unlearning corpus.
  • domain assumption CLIP cosine similarity between logo region and company-name text measures brand recognizability.
    Defines CLIPScore in §3.3.2; no calibration against human perception.
  • domain assumption stable-diffusion-3-medium is representative and generates logos reliably after prompt filtering.
    Main backbone (§3.1); generalization to other modern T2I models is asserted via Figure 3 but not quantified.
  • domain assumption Human reviewers' filtering yields natural, valid prompts and correct implicit cues.
    Dataset construction (§3.1); reviewer count and agreement are not reported.
  • domain assumption SSIM and CLIP similarities are appropriate for quantifying structural and semantic preservation.
    Used for LogoSSIM/ImageSSIM and LogoScore/ImageScore; standard metrics but no validation for this task.

pith-pipeline@v1.3.0-alltime-deepseek · 14485 in / 16109 out tokens · 136510 ms · 2026-07-31T23:03:31.949615+00:00 · methodology

0 comments
read the original abstract

Concept unlearning is increasingly used to limit the reproduction of protected or unsafe visual concepts in text-to-image models. Existing evaluations, however, mostly study targets that dominate the whole image, such as styles, broad object categories, or portrait-like identities, leaving company logos comparatively underexamined. Logos create a different failure mode: a small localized mark can carry the entire protected concept, must be visually precise to remain recognizable, and can be triggered implicitly by products, storefronts, packaging, or advertisements even when the word ``logo'' is absent. We introduce LU-500, a logo-unlearning benchmark built from Fortune Global 500 companies to study this localized and semantically entangled setting. LU-500 contains nearly 10,000 curated text-query and logo-image pairs, with an explicit track (LUex-500) and an implicit contextual track (LUim-500). To avoid reducing the task to a binary detector score, we define a multi-grained protocol that evaluates both local logo removal and global image preservation in pixel and latent spaces. Experiments on representative inference-time methods, including NP, SLD, and SEGA, and compatible fine-tuning-based methods such as ESD and Forget-Me-Not, show that the evaluated methods struggle to remove logo evidence without changing non-target content. We further analyze ProLU, a prompt-space multi-agent baseline: it improves local erasure by removing logo-inducing semantics, but also illustrates why prompt filtering is not a substitute for weight-level disentanglement. Correlation analyses over logo area, location, and structural complexity suggest that future logo unlearning may need spatially aware controls, such as SSIM-guided constraints, rather than purely global concept suppression.

Figures

Figures reproduced from arXiv: 2607.24101 by Dequan Wang, Jialing Zhang, Jin Gao, Keyu Li.

Figure 1
Figure 1. Figure 1: Overview of LU-500, a benchmark evaluating logo unlearning on the Fortune Global 500. We propose multi-grained metrics assessing local concept erasure and global image fidelity. Evaluations reveal existing inference-time approaches (e.g., NP [1], SLD [2], SEGA [3]) systemati￾cally fail to unlearn these entangled assets. Our prompt-based baseline, ProLU, achieves superior local erasure but exposes a critica… view at source ↗
Figure 2
Figure 2. Figure 2: Prompt distributions across LU-500 tracks and their evolutionary trajectories during semantic unlearning. P0 is the initial logo-inducing prompt. P1 and P1 ∗ denote intermediate states during iterative concept ablation by the ProLU Reflector agent. P2 is the final sanitized prompt for T2I generation. These statistical profiles highlight the explicit logo references of the Easy track versus the implicit con… view at source ↗
Figure 3
Figure 3. Figure 3: Escalating high-fidelity logo leakage in modern T2I architectures. While legacy models (SD1.5) struggle with complex brand iconography, contemporary models (SD3, SD3.5, FLUX) precisely hallucinate the copyrighted “FedEx” logo. This capability jump necessitates LU-500 and justifies selecting SD3 to evaluate concept unlearning. models: can a method prevent or remove a localized brand mark while preserving th… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the LU-500 pipeline and ProLU exploratory framework. Left: The human-AI collaborative generation of LU-500. Using Apple as an example, it details explicit and implicit prompt synthesis via GPT-5, multi-stage human validation, and the multi-grained evaluation protocol. Right: The semantic intervention logic of ProLU. It orchestrates the Remover, Reflector, and Checker agents to iteratively ablat… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of inference-time unlearning methods across diverse logo topologies. Samples span pure geometric icons (e.g., Apple, Tesla), typography-centric designs (e.g., Coca-Cola, Walt Disney), and complex composites (e.g., Mercedes-Benz). These results visually corroborate the systemic trade-off in latent-space interventions: scaling safety guidance (left to right for SLD/SEGA) distorts the l… view at source ↗
Figure 6
Figure 6. Figure 6: Diagnostic correlation between visual attributes and SEGA [3] unlearning efficacy. We compute Pearson correlations across intrinsic geometric and textural properties (e.g., area, texture complexity). While most factors exhibit negligible correlation, LogoSSIM (local SSIM) reveals a notable relationship. This suggests that incorporating structural heuristics (e.g., SSIM-guided spatial controls) is vital for… view at source ↗
Figure 7
Figure 7. Figure 7: Original images and their unlearned counterparts are analyzed using metrics focused on [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The agent prompts for generating LU-500 was crafted using OpenAI’s GPT-4o model. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The Remover, Reflector, and Checker prompts in ProLU. The Remover is used to eliminate [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: More visual results over Boeing, DELL, EXXON MOBIL, FedEx, Goldman-Sachs, HP, [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: More visual results on Oracle, Pfizer, Qualcomm, Renault, Uber, Volvo, Walt Disney, [PITH_FULL_IMAGE:figures/full_fig_p021_11.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

66 extracted references · 17 linked inside Pith

  1. [1]

    Stable diffusion 2.0 release, 2023

    Dana Rao. Stable diffusion 2.0 release, 2023. URL https://stability.ai/news/ stable-diffusion-v2-release

    2023

  2. [2]

    Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models

    Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522–22531, 2023

    2023

  3. [3]

    Sega: Instructing text-to-image models using semantic guidance.Advances in Neural Information Processing Systems, 36:25365–25389, 2023

    Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance.Advances in Neural Information Processing Systems, 36:25365–25389, 2023

    2023

  4. [4]

    Unlearning concepts in diffusion model via concept domain correction and concept preserving gradient.arXiv preprint arXiv:2405.15304, 2024

    Yongliang Wu, Shiji Zhou, Mingzhuo Yang, Lianzhe Wang, Wenbo Zhu, Heng Chang, Xiao Zhou, and Xu Yang. Unlearning concepts in diffusion model via concept domain correction and concept preserving gradient.arXiv preprint arXiv:2405.15304, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Meta- unlearning on diffusion models: Preventing relearning unlearned concepts.arXiv preprint arXiv:2410.12777, 2024

    Hongcheng Gao, Tianyu Pang, Chao Du, Taihang Hu, Zhijie Deng, and Min Lin. Meta- unlearning on diffusion models: Preventing relearning unlearned concepts.arXiv preprint arXiv:2410.12777, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    A survey of machine unlearning.arXiv preprint arXiv:2209.02299, 2022

    Thanh Tam Nguyen, Thanh Trung Huynh, Phi Le Nguyen, Alan Wee-Chung Liew, Hongzhi Yin, and Quoc Viet Hung Nguyen. A survey of machine unlearning.arXiv preprint arXiv:2209.02299, 2022

    Pith/arXiv arXiv 2022

  7. [7]

    Avoiding copyright infringement via machine unlearning.arXiv preprint arXiv:2406.10952, 2024

    Guangyao Dou, Zheyuan Liu, Qing Lyu, Kaize Ding, and Eric Wong. Avoiding copyright infringement via machine unlearning.arXiv preprint arXiv:2406.10952, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Shield: Evaluation and defense strategies for copyright compliance in llm text generation.arXiv preprint arXiv:2406.12975, 2024

    Xiaoze Liu, Ting Sun, Tianyang Xu, Feijie Wu, Cunxiang Wang, Xiaoqian Wang, and Jing Gao. Shield: Evaluation and defense strategies for copyright compliance in llm text generation.arXiv preprint arXiv:2406.12975, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Copyright traps for large language models.arXiv preprint arXiv:2402.09363, 2024

    Matthieu Meeus, Igor Shilov, Manuel Faysse, and Yves-Alexandre de Montjoye. Copyright traps for large language models.arXiv preprint arXiv:2402.09363, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Six-cd: Benchmarking concept removals for benign text-to-image diffusion models.arXiv preprint arXiv:2406.14855, 2024

    Jie Ren, Kangrui Chen, Yingqian Cui, Shenglai Zeng, Hui Liu, Yue Xing, Jiliang Tang, and Lingjuan Lyu. Six-cd: Benchmarking concept removals for benign text-to-image diffusion models.arXiv preprint arXiv:2406.14855, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    stable-diffusion-3-medium, 2024

    stabilityai. stable-diffusion-3-medium, 2024. URL https://huggingface.co/ stabilityai/stable-diffusion-3-medium

    2024

  12. [12]

    Scaling rectified flow transform- ers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transform- ers for high-resolution image synthesis. InForty-first International Conference on Machine Learning, 2024

    2024

  13. [13]

    High- resolution image synthesis with latent diffusion models

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

    2022

  14. [14]

    Flux1.1 pro, 2024

    Black Forest Labs. Flux1.1 pro, 2024. URLhttps://blackforestlabs.ai/

    2024

  15. [15]

    Erasing concepts from diffusion models

    Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426–2436, 2023

    2023

  16. [16]

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

    Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me- not: Learning to forget in text-to-image diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1755–1764, 2024

    2024

  17. [17]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4): 600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4): 600–612, 2004. 11

    2004

  18. [18]

    Logo-2k+: A large-scale logo dataset for scalable logo classification

    Jing Wang, Weiqing Min, Sujuan Hou, Shengnan Ma, Yuanjie Zheng, Haishuai Wang, and Shuqiang Jiang. Logo-2k+: A large-scale logo dataset for scalable logo classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6194–6201, 2020

    2020

  19. [19]

    Weblogo-2m: Scalable logo detection by deep learning from the web

    Hang Su, Shaogang Gong, and Xiatian Zhu. Weblogo-2m: Scalable logo detection by deep learning from the web. InProceedings of the IEEE international conference on computer vision workshops, pages 270–279, 2017

    2017

  20. [20]

    Logo retrieval with a contrario visual query expansion

    Alexis Joly and Olivier Buisson. Logo retrieval with a contrario visual query expansion. In Proceedings of the 17th ACM international conference on Multimedia, pages 581–584, 2009

    2009

  21. [21]

    Scalable logo recognition in real-world images

    Stefan Romberg, Lluis Garcia Pueyo, Rainer Lienhart, and Roelof Van Zwol. Scalable logo recognition in real-world images. InProceedings of the 1st ACM international conference on multimedia retrieval, pages 1–8, 2011

    2011

  22. [22]

    A deep one-shot network for query-based logo retrieval.Pattern Recognition, 96:106965, 2019

    Ayan Kumar Bhunia, Ankan Kumar Bhunia, Shuvozit Ghose, Abhirup Das, Partha Pratim Roy, and Umapada Pal. A deep one-shot network for query-based logo retrieval.Pattern Recognition, 96:106965, 2019

    2019

  23. [23]

    The open brands dataset: Unified brand detection and recognition at scale

    Xuan Jin, Wei Su, Rong Zhang, Yuan He, and Hui Xue. The open brands dataset: Unified brand detection and recognition at scale. InICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4387–4391. IEEE, 2020

    2020

  24. [24]

    Multi-scale vehicle logo detector

    Junxing Zhang, Lijun Chen, Chunjuan Bo, and Shuo Yang. Multi-scale vehicle logo detector. Mobile Networks and Applications, 26:67–76, 2021

    2021

  25. [25]

    Foodlogodet-1500: A dataset for large-scale food logo detection via multi-scale feature decou- pling network

    Qiang Hou, Weiqing Min, Jing Wang, Sujuan Hou, Yuanjie Zheng, and Shuqiang Jiang. Foodlogodet-1500: A dataset for large-scale food logo detection via multi-scale feature decou- pling network. InProceedings of the 29th ACM International Conference on Multimedia, pages 4670–4679, 2021

    2021

  26. [26]

    Seetek: Very large-scale open-set logo recognition with text-aware metric learning

    Chenge Li, István Fehérvári, Xiaonan Zhao, Ives Macedo, and Srikar Appalaraju. Seetek: Very large-scale open-set logo recognition with text-aware metric learning. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2544–2553, 2022

    2022

  27. [27]

    Deep learning for logo detection: A survey.ACM Transactions on Multimedia Computing, Communications and Applications, 20(3):1–23, 2023

    Sujuan Hou, Jiacheng Li, Weiqing Min, Qiang Hou, Yanna Zhao, Yuanjie Zheng, and Shuqiang Jiang. Deep learning for logo detection: A survey.ACM Transactions on Multimedia Computing, Communications and Applications, 20(3):1–23, 2023

    2023

  28. [28]

    Logodet-3k: A large-scale image dataset for logo detection.ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 18(1):1–19, 2022

    Jing Wang, Weiqing Min, Sujuan Hou, Shengnan Ma, Yuanjie Zheng, and Shuqiang Jiang. Logodet-3k: A large-scale image dataset for logo detection.ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 18(1):1–19, 2022

    2022

  29. [29]

    Open logo detection challenge.arXiv preprint arXiv:1807.01964, 2018

    Hang Su, Xiatian Zhu, and Shaogang Gong. Open logo detection challenge.arXiv preprint arXiv:1807.01964, 2018

    Pith/arXiv arXiv 2018

  30. [30]

    Logo-net: Large-scale deep logo detection and brand recognition with deep region-based convolutional networks.arXiv preprint arXiv:1511.02462, 2015

    Steven CH Hoi, Xiongwei Wu, Hantang Liu, Yue Wu, Huiqiong Wang, Hui Xue, and Qiang Wu. Logo-net: Large-scale deep logo detection and brand recognition with deep region-based convolutional networks.arXiv preprint arXiv:1511.02462, 2015

    Pith/arXiv arXiv 2015

  31. [31]

    Open set logo detection and retrieval.arXiv preprint arXiv:1710.10891, 2017

    Andras Tüzkö, Christian Herrmann, Daniel Manger, and Jürgen Beyerer. Open set logo detection and retrieval.arXiv preprint arXiv:1710.10891, 2017

    Pith/arXiv arXiv 2017

  32. [32]

    Comparative analysis of deep learning models for brand logo classification in real-world scenarios.arXiv preprint arXiv:2305.12242, 2023

    Qimao Yang, Huili Chen, and Qiwei Dong. Comparative analysis of deep learning models for brand logo classification in real-world scenarios.arXiv preprint arXiv:2305.12242, 2023

    Pith/arXiv arXiv 2023

  33. [33]

    Aligned agents, biased swarm: Measuring bias amplifica- tion in multi-agent systems.arXiv preprint arXiv:2604.08963, 2026

    Keyu Li, Jin Gao, and Dequan Wang. Aligned agents, biased swarm: Measuring bias amplifica- tion in multi-agent systems.arXiv preprint arXiv:2604.08963, 2026

    Pith/arXiv arXiv 2026

  34. [34]

    davinci-agency: Unlocking long-horizon agency data-efficiently

    Mohan Jiang, Dayuan Fu, Junhao Shi, Ji Zeng, Weiye Si, Keyu Li, Xuefeng Li, Yang Xiao, Wenjie Li, Dequan Wang, et al. davinci-agency: Unlocking long-horizon agency data-efficiently. arXiv preprint arXiv:2602.02619, 2026. 12

    arXiv 2026

  35. [35]

    Kan-mixer: Kolmogorov-arnold networks for gene expression prediction in plant species

    Jin Gao, Juntu Zhao, Keyu Li, and Dequan Wang. Kan-mixer: Kolmogorov-arnold networks for gene expression prediction in plant species. InEuropean Conference on Computer Vision, pages 135–150. Springer, 2024

    2024

  36. [36]

    Towards safe self-distillation of internet-scale text-to-image diffusion models.arXiv preprint arXiv:2307.05977, 2023

    Sanghyun Kim, Seohyeon Jung, Balhae Kim, Moonseok Choi, Jinwoo Shin, and Juho Lee. Towards safe self-distillation of internet-scale text-to-image diffusion models.arXiv preprint arXiv:2307.05977, 2023

    Pith/arXiv arXiv 2023

  37. [37]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559–7568, 2024

    2024

  38. [38]

    Unified concept editing in diffusion models

    Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzy´nska, and David Bau. Unified concept editing in diffusion models. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111–5120, 2024

    2024

  39. [39]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430–6440, 2024

    2024

  40. [40]

    Editing massive concepts in text-to-image diffusion models.arXiv preprint arXiv:2403.13807, 2024

    Tianwei Xiong, Yue Wu, Enze Xie, Zhenguo Li, and Xihui Liu. Editing massive concepts in text-to-image diffusion models.arXiv preprint arXiv:2403.13807, 2024

    arXiv 2024

  41. [41]

    A dataset and benchmark for copyright protection from text-to-image diffusion models.arXiv preprint arXiv:2403.12052, 2024

    Rui Ma, Qiang Zhou, Bangjun Xiao, Yizhu Jin, Daquan Zhou, Xiuyu Li, Aishani Singh, Yi Qu, Kurt Keutzer, Xiaodong Xie, et al. A dataset and benchmark for copyright protection from text-to-image diffusion models.arXiv preprint arXiv:2403.12052, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models.arXiv preprint arXiv:2402.11846, 2024

    Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models.arXiv preprint arXiv:2402.11846, 2024

    Pith/arXiv arXiv 2024

  43. [43]

    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. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691–22702, 2023

    2023

  44. [44]

    Limi: Less is more for agency.arXiv preprint arXiv:2509.17567, 2025

    Yang Xiao, Mohan Jiang, Jie Sun, Keyu Li, Jifan Lin, Yumin Zhuang, Ji Zeng, Shijie Xia, Qishuo Hua, Xuefeng Li, et al. Limi: Less is more for agency.arXiv preprint arXiv:2509.17567, 2025

    arXiv 2025

  45. [45]

    Datasetresearch: Benchmarking agent systems for demand-driven dataset discovery.arXiv preprint arXiv:2508.06960, 2025

    Keyu Li, Mohan Jiang, Dayuan Fu, Yunze Wu, Xiangkun Hu, Dequan Wang, and Pengfei Liu. Datasetresearch: Benchmarking agent systems for demand-driven dataset discovery.arXiv preprint arXiv:2508.06960, 2025

    Pith/arXiv arXiv 2025

  46. [46]

    Agencybench: Benchmarking the frontiers of autonomous agents in 1m-token real-world contexts.arXiv preprint arXiv:2601.11044, 2026

    Keyu Li, Junhao Shi, Yang Xiao, Mohan Jiang, Jie Sun, Yunze Wu, Shijie Xia, Xiaojie Cai, Tianze Xu, Weiye Si, et al. Agencybench: Benchmarking the frontiers of autonomous agents in 1m-token real-world contexts.arXiv preprint arXiv:2601.11044, 2026

    Pith/arXiv arXiv 2026

  47. [47]

    Innovatorbench: Evaluating agents’ ability to conduct innovative llm research.arXiv preprint arXiv:2510.27598, 2025

    Yunze Wu, Dayuan Fu, Weiye Si, Zhen Huang, Mohan Jiang, Keyu Li, Shijie Xia, Jie Sun, Tianze Xu, Xiangkun Hu, et al. Innovatorbench: Evaluating agents’ ability to conduct innovative llm research.arXiv preprint arXiv:2510.27598, 2025

    arXiv 2025

  48. [48]

    Midjourney (V5.2) [Text-to-Image Model], 2023

    Midjourney. Midjourney (V5.2) [Text-to-Image Model], 2023. URL https://www. midjourney.com

    2023

  49. [49]

    Dalle3, 2024

    openai. Dalle3, 2024. URLhttps://openai.com/index/dall-e-3/

    2024

  50. [50]

    Scaling open-vocabulary object detection.Advances in Neural Information Processing Systems, 36, 2024

    Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection.Advances in Neural Information Processing Systems, 36, 2024

    2024

  51. [51]

    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. InInternational conference on machine learning, pages 8748–8763. PMLR, 2021. 13

    2021

  52. [52]

    stable-diffusion-v1-5, 2022

    runawayml. stable-diffusion-v1-5, 2022. URL https://huggingface.co/runwayml/ stable-diffusion-v1-5

    2022

  53. [53]

    A computational approach to edge detection, 1986

    John Canny. A computational approach to edge detection, 1986

    1986

  54. [54]

    Texture characterization based on grey-level co-occurrence matrix

    A Gebejes and R Huertas. Texture characterization based on grey-level co-occurrence matrix. Databases, 9(10):375–378, 2013

    2013

  55. [55]

    Divergence measures based on the shannon entropy.IEEE Transactions on Information theory, 37(1):145–151, 1991

    Jianhua Lin. Divergence measures based on the shannon entropy.IEEE Transactions on Information theory, 37(1):145–151, 1991

    1991

  56. [56]

    apple logo

    Karl Pearson. Vii. mathematical contributions to the theory of evolution.—iii. regression, heredity, and panmixia, 1896. 14 A Comprehensive Evaluation Metrics We establish a comprehensive, multi-grained evaluation framework comprising five distinct metrics, each with a stringently defined scope to capture the complex trade-offs between local concept erasu...

  57. [57]

    A photo of a closed MacBook with an Apple logo

  58. [58]

    A photo of an iPad with an Apple logo

  59. [59]

    A photo of an iPhone with an Apple logo

  60. [60]

    Requirements: Generate 10 prompts directly in English, format them as follows:

    A photo of an Apple store with an Apple logo in front of it. Requirements: Generate 10 prompts directly in English, format them as follows:

  61. [61]

    Keep the prompts simple, without complex scenes

    prompt... Keep the prompts simple, without complex scenes. Please generate ten text-to-image prompts for {company} that will result in images featuring the {company}'s logo prominently. Consider emphasizing the company's founder, famous products, competitors, company's products, stores, advertisements, office buildings, employees, or website, so that the ...

  62. [62]

    A photo of a closed MacBook

  63. [63]

    A photo of an iPad in use

  64. [64]

    A photo of the back of an iPhone

  65. [65]

    Make the prompt as detailed as possible, but not overly lengthy

    A photo of an Apple store entrance. Make the prompt as detailed as possible, but not overly lengthy. It may include complex scenes except for the company logo. You can add other scenes to add complexity in prompts. You may mention the company name in the prompt to ensure the resulting images will feature the logo prominently. Requirements: Generate 10 pro...

  66. [66]

    Explicit Implicit Figure 8: The agent prompts for generating LU-500 was crafted using OpenAI’s GPT-4o model

    prompt ... Explicit Implicit Figure 8: The agent prompts for generating LU-500 was crafted using OpenAI’s GPT-4o model. 18 Figure 9: The Remover, Reflector, and Checker prompts in ProLU. The Remover is used to eliminate elements related to the company logo from the original prompt while keeping other parts as consistent as possible. The Reflector evaluate...