Pith. sign in

REVIEW 4 major objections 6 minor 48 references

How Robust is Model Editing after Fine-Tuning? An Empirical Study on Text-to-Image Diffusion Models

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning erases prior edits in image-generation models

desk verdict A broad empirical study of edit persistence after fine-tuning, but the gender-debiasing metric is unsigned and undermines the headline claim, so the paper needs major revision before the conclusion is trustworthy. read the letter →

arxiv 2506.18428 v1 pith:S6H7FBNQ submitted 2025-06-23 cs.AI cs.LG

Feng He , Zhenyang Liu , Marco Valentino , Zhixue Zhao This is my paper · ORCID
classification cs.AIcs.LG
keywords modeleditingtext-to-imagediffusionfine-tuningeditpersistencegenderdebiasingunsafecontentremovalparameter-efficientconcept
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 asks whether a behavior injected into a text-to-image model by model editing survives a later fine-tuning step. Across two editing methods, three model families, and four fine-tuning approaches, the authors report that it generally does not: fine-tuning degrades or reverses edits even when the fine-tuning task is stylistically unrelated to the edited behavior. The finding matters because deployed models are often edited for bias mitigation or unsafe-content removal and then fine-tuned for style or domain adaptation. If the paper is right, beneficial edits must be re-applied after fine-tuning, and fine-tuning becomes a plausible way to scrub hidden malicious edits.

What carries the argument

The load-bearing setup is a four-way model comparison: $M$ (original), $M_{ed}$ (edited), $M_{ft}$ (fine-tuned), and $M_{ed\text{-}ft}$ (edited then fine-tuned). Persistence is quantified as $\Delta(\psi; M_{ed}, M_{ed\text{-}ft})$, the normalized behavioral discrepancy on prompts relevant to the edit $\psi$; a small $\Delta$ means the edit survived. To measure it, the paper uses CLIP-based concept similarity, the gender-ratio deviation $\delta$, human safety annotations on I2P prompts, FID, and CLIP Score. The candidate explanation for reversal is polysemantic neurons, units that encode several unrelated concepts at once, whose stored version of the unwanted concept can be re-activated while fine-tuning strengthens other features.

What would settle it

Rerun the gender-debiasing and unsafe-removal evaluations on the edited-then-fine-tuned models using human annotators or a style-invariant classifier trained on anime-style images; if the edited behaviors are found intact under those measures, the reversal claim is a measurement artifact rather than a genuine loss.

Watch

Extended reading notes

Core claim

The central claim is that the cascade edit-then-fine-tune collapses the edit: an edited model that is later fine-tuned behaves more like a model that was only fine-tuned than like the edited model. The evidence comes from a four-way comparison of base, edited, fine-tuned, and edited-then-fine-tuned variants of Stable Diffusion v1.4, SDXL, and FLUX. Across gender-debiasing, appearance, role, and unsafe-concept-removal tasks, fine-tuning weakens the edit; full-size fine-tuning and DoRA produce the largest reversal, DreamBooth the smallest, and UCE-based edits survive noticeably better than ReFACT-based edits. The authors conclude that editing alone is not durable through the deployment lifecycle, which makes fine-tuning both a potential defense against malicious edits and a hazard to beneficial safety and alignment edits.

Load-bearing premise

The metrics used to decide whether an edit is still present remain valid after fine-tuning shifts the image style, so the reported reversals reflect genuine loss of the edit rather than artifacts of scoring anime-style images with CLIP-based classifiers.

Editorial extensions

If this is right

  • Safety edits that remove nudity or violence can be partially undone by later style fine-tuning, so deployments should re-audit safety after any fine-tuning.
  • Gender-debiasing edits degrade across all tested fine-tuning methods, so maintaining parity after adaptation requires re-editing.
  • Full-size fine-tuning and DoRA are the most effective at erasing prior edits, making them the practical choice for remediating hidden or malicious edits.
  • DreamBooth preserves edits best and LoRA is more stable than DoRA, so lightweight tuning methods are preferable when edit retention is the priority; among editors, UCE is more durable than ReFACT.

Reading between the lines

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

  • A testable extension is to vary how semantically close the fine-tuning data is to the edited concept: the reversal should be stronger when the two overlap, and weaker when they are entirely disjoint.
  • Editing methods could be made lifecycle-aware by adding a regularization term that penalizes drift under simulated fine-tuning updates, turning edit persistence into an explicit optimization objective.
  • Part of the observed reversal may be a measurement artifact, since CLIP-based gender and concept classifiers are validated on photorealistic images but applied to anime-style outputs; human re-annotation on style-shifted images would separate true edit loss from metric bias.
  • The results suggest treating edit persistence as a standard evaluation dimension for editing methods, analogous to how continual learning tracks forgetting, rather than evaluating edits in isolation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper asks whether concept edits applied to text-to-image diffusion models survive a subsequent fine-tuning step, and it reports an empirical study spanning two editing methods (UCE, ReFACT), three model families (SD1.4, SDXL, FLUX), and four fine-tuning approaches (DreamBooth, full fine-tuning, LoRA, DoRA). The authors measure edit persistence using gender-debiasing statistics (δ), appearance/role efficacy, unsafe-content annotation counts, and FID/CLIP scores. Their central conclusion is that edits generally fail to persist through fine-tuning, that DoRA reverses edits most aggressively, and that UCE is more robust than ReFACT. The paper also draws safety implications: fine-tuning could remediate malicious edits, while beneficial edits may need to be reapplied after fine-tuning.

Significance. If the central claim is correct, the paper addresses an important and understudied question in the model-editing deployment lifecycle: whether edited behavior survives subsequent adaptation. The experimental breadth—multiple editing tasks, model families, and fine-tuning methods—is a genuine strength, as is the inclusion of human annotation for unsafe content with reported inter-annotator agreement. The paper is also largely reproducible in principle, since it uses public datasets and open-source methods. However, the quantitative evidence as presented is internally inconsistent in several load-bearing places: the key gender-debiasing table uses an unsigned metric that mixes improvements with degradations, the DoRA 'strongest reversal' claim is not supported by the reported averages, and the unsafe-content tables pool opposite directions of label flips. These issues mean the headline claims are not currently established by the data, although they may be restorable with a corrected, signed analysis.

major comments (4)
  1. [4.2, Table 1] The claim that 'all four fine-tuning methods lead to a degradation of the gender debiasing effect across all six professions' is contradicted by the paper's own table. Δ is defined as the absolute difference |δ(Med) − δ(Med_ft)|, so a move toward parity is scored identically to a move away from parity. Reading Table 1 with signs, at least 11 of the 24 cells move toward parity (e.g., Teacher+LoRA: δ goes 0.53→0.24; Housekeeper+Full: 0.58→0.43; Housekeeper+DoRA: 0.58→0.43), while the remaining cells move away. A signed count gives roughly 13 away-from-parity and 11 toward-parity changes. The conclusion that edits 'generally fail to persist' is therefore not established by the gender-debiasing experiment as reported. Please report signed changes and per-profession significance tests, and recompute any aggregate claims.
  2. [Abstract; 4.2, Table 1; 4.4, Table 3] The claim that 'DoRA exhibits the strongest edit reversal effect' is unsupported by the reported data. In Table 1, full-size fine-tuning, DoRA, and LoRA all have the same average absolute Δ of 0.115, with only DreamBooth lower (0.057). In Table 3, the unsafe-content results are also mixed: full-size fine-tuning increases the number of safe images from 36 in Med to 39 in Med_ft, and DoRA increases it from 36 to 42, while LoRA decreases it to 31. No single metric or aggregate in the paper shows DoRA as the strongest reversal across tasks. If the claim is based on other results (e.g., Table 2 variance or qualitative step analysis), please specify the exact evidence and the chosen metric.
  3. [3.4, 4.4, Table 3] The unsafe-content analysis pools opposite directions of label changes. The Flip score counts any label change regardless of direction, and Δ is an unsigned net change, so a safe→unsafe flip and an unsafe→safe flip are treated as equally damaging. In Table 3, for full-size fine-tuning, safe images increase from 36 to 39 and black images drop from 7 to 0, which is not obviously 'weakening the safety level' as claimed. Similarly, the FLUX result in Table 4 reports 12 flips away from safe but does not report how many unsafe/can't-decide images became safe. Please report signed transitions (e.g., a confusion matrix between Med and Med_ft labels) and define which transitions count as degradation.
  4. [3.4, 4.2] The evaluation metrics are applied across a large style shift without validation. CLIP-based gender classification and concept-similarity scores are used on anime-style images produced after Naruto-style fine-tuning, but the edits and baselines are validated on photorealistic outputs. If the CLIP classifier's gender decisions or the ReFACT similarity measures are biased by the style shift, the observed 'reversal' could be partly a measurement artifact. The paper should either validate the metrics on style-shifted images (e.g., human-annotated gender labels on a sample of Med_ft images, or per-style classifier accuracy) or explicitly report sensitivity of the main conclusions to this shift.
minor comments (6)
  1. [3.2, 3.3] There are typos: 'supplimentary' should be 'supplementary' (Section 3.2), and 'attampt' should be 'attempt' (Section 3.3).
  2. [3.4, 10.1] The definition of δ in Section 3.4 relies on F_p, the percentage of generated female-presenting images, but F_p is only defined explicitly in Section 10.1. Please define F_p at first use.
  3. [Figure 3] Figure 3 presents efficacy, generality, and specificity differences only as bar charts without numeric values or error bars; adding the underlying numbers would make the claimed degradation pattern in Section 4.3 verifiable.
  4. [Table 1] The 'Avg.(std.)' row is formatted confusingly: it mixes standard deviations of δ values with average Δ values, and the standard deviations are reported without explaining that they summarize per-profession δ variation. Please separate the summary statistics for δ and Δ.
  5. [4.2] The explanation that DoRA overrides edits because of Prodigy's aggressive learning rate schedule is presented as a conclusion but is only a hypothesis. Please label it as speculative or provide supporting evidence.
  6. [References] References [40] and [41] are the same paper (Safe Latent Diffusion); please merge them.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the claimed edit-persistence results are direct empirical measurements, not quantities that reduce to their inputs by construction.

full rationale

This is an empirical measurement study, not a derivation. The central claim that 'edits generally fail to persist through fine-tuning' is supported by direct comparisons between the edited model Med and the edited-then-fine-tuned model Med_ft, using the discrepancy defined in Eq. (4) as Delta = |E[R(Med)] - E[R(Med_ft)]|. Delta is a measured difference between two independently produced model variants; it is not a fitted parameter, nor is it defined in terms of the paper's conclusion. The evaluation protocols (RoAD for ReFACT, gender-ratio deviation on WinoBias-style professions, I2P human annotation for unsafe content, and FID/CLIP for quality) are external benchmarks, and the editing and fine-tuning methods are adopted from prior work with their recommended hyperparameters. The only self-citation in the author set is reference [46], a position paper on LLM editing risks; it is cited only to support the premise that harmful edits exist, not as evidence for the empirical persistence findings, so it is not load-bearing. The paper's own Limitations section concedes limited dataset diversity and qualitative examples, which bears on generality and correctness risk rather than circularity. Potential methodological concerns, such as style shift influencing CLIP-based gender classification or the unsigned Delta conflating moves toward and away from parity, are validity questions about the metric; they do not make the claim equivalent to its inputs by construction. No load-bearing step reduces to a fit, a self-citation, or a definitional equivalence.

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

The central empirical claims rest on the validity of the evaluation metrics and the representativeness of the fine-tuning datasets. No free parameters or invented entities are introduced.

assumptions (4)
  • domain assumption CLIP-based gender classification is accurate across the generated image styles used in the study, including anime style after fine-tuning.
    Section 3.4 uses CLIP to compute female ratio delta for profession images; if style shift biases CLIP, the debiasing results are confounded.
  • domain assumption Human annotations of unsafe content are reliable and consistent.
    Section 3.4 and 9.2: human evaluation on 50 I2P prompts; Fleiss' kappa 0.717 supports this, though it is based on a small pilot.
  • domain assumption The Naruto-style and DreamBooth datasets are representative of typical downstream fine-tuning.
    Section 3.2: only two small datasets are used (1,220 images and 30 subjects); the paper acknowledges this limitation in Section 8.
  • domain assumption The metrics efficacy, generality, and specificity correctly capture edit persistence.
    Section 3.4: these metrics are adopted from prior work without validating their sensitivity to fine-tuning-induced style shift.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Robust is Model Editing after Fine-Tuning? An Empirical Study on Text-to-Image Diffusion Models." pith.science (2026). https://pith.science/paper/S6H7FBNQ

@misc{pith2026250618428,
  author       = {Pith},
  title        = {Pith review of: How Robust is Model Editing after Fine-Tuning? An Empirical Study on Text-to-Image Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S6H7FBNQ}},
  note         = {Machine review of arXiv:2506.18428}
}
read the original abstract

Model editing offers a low-cost technique to inject or correct a particular behavior in a pre-trained model without extensive retraining, supporting applications such as factual correction and bias mitigation. Despite this common practice, it remains unknown whether edits persist after fine-tuning or whether they are inadvertently reversed. This question has fundamental practical implications. For example, if fine-tuning removes prior edits, it could serve as a defence mechanism against hidden malicious edits. Vice versa, the unintended removal of edits related to bias mitigation could pose serious safety concerns. We systematically investigate the interaction between model editing and fine-tuning in the context of T2I diffusion models, which are known to exhibit biases and generate inappropriate content. Our study spans two T2I model families (Stable Diffusion and FLUX), two sota editing techniques, and three fine-tuning methods (DreamBooth, LoRA, and DoRA). Through an extensive empirical analysis across diverse editing tasks and evaluation metrics, our findings reveal a trend: edits generally fail to persist through fine-tuning, even when fine-tuning is tangential or unrelated to the edits. Notably, we observe that DoRA exhibits the strongest edit reversal effect. At the same time, among editing methods, UCE demonstrates greater robustness, retaining significantly higher efficacy post-fine-tuning compared to ReFACT. These findings highlight a crucial limitation in current editing methodologies, emphasizing the need for more robust techniques to ensure reliable long-term control and alignment of deployed AI systems. These findings have dual implications for AI safety: they suggest that fine-tuning could serve as a remediation mechanism for malicious edits while simultaneously highlighting the need for re-editing after fine-tuning to maintain beneficial safety and alignment properties.

Figures

Figures reproduced from arXiv: 2506.18428 by the authors.

Figure 1
Figure 1. Overview of our workflow (left). Starting from the original model [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Images generated by original SD1.4, DoRA-fine-tuned SD1.4, LoRA-fine-tuned SD1.4, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Efficacy, Generality, and Specificity on SD1.4. The light grey and dark grey bars represent [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: The example of editing appearances and roles using ReFACT. The left shows an appearance [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Impact of fine-tuning on the edited “unsafe” concept over training steps in SD1.4. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Examples of applying UCE to SDXL. After editing, the model frequently generates chaotic [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Overview of M, Mft, Med (UCE debias), and Med_ft. Fine-tuning methods (top to bottom): DreamBooth, full-size finetuning, DoRA, and LoRA. Visualization of Appearance and Role Edit Performance Edit performance of four models (M, Med, Mft, Med_ft) is visualized using fill…
Figure 8
Figure 8. Figure 8: Comparison of M, Med (ReFACT-Role), and Mft (full size, LoRA, DoRA and Dream￾Booth). Appearance Full Size LoRA DoRA DB Edit “computer” to “laptop” Prompt: “A computer and a plant on a workstation” Edit “telephone” to “smartphone” Prompt: “A telephone on a table” M Med …
Figure 9
Figure 9. Figure 9: Comparison of M, Med (ReFACT-Appearance), and Mft. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Efficacy, Generality, and Specificity on SDXL. The light grey and dark grey bars represent [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Overview of applying UCE debiasing followed by full-size fine-tuning. [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Overview of applying UCE debiasing followed by DoRA. [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Overview of applying UCE debiasing followed by LoRA. [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: UCE erasure task (violence) with M, Med, and three Med_ft variants. 10.4 Generation Quality: Clip Score and FID We evaluate image quality using FID [17] and CLIP Score [16] across the base model and its three variants. Overall, we find that Med preserves generation qu…
Figure 15
Figure 15. Figure 15: UCE erasure task (nudity) with M, Med, and three Med_ft variants. generated images to deviate from MS COCO, a real-world dataset used as the reference for FID computation. Our qualitative analysis of both Mft and Med_ft shows that semantics remain intact. DreamBooth, …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 21 canonical work pages

  1. [1]

    ReFACT: Updating text-to-image models by editing the text encoder

    Dana Arad, Hadas Orgad, and Yonatan Belinkov. ReFACT: Updating text-to-image models by editing the text encoder. In Kevin Duh, Helena Gomez, and Steven Bethard, editors, Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (V olume1: Long Papers), pages 2537–2558, Me...

    doi:10.18653/v1/2024.naacl-long.140 2024

  2. [2]

    Easily accessible text-to- image generation amplifies demographic stereotypes at large scale

    Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to- image generation amplifies demographic stereotypes at large scale. In 2023 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’23, page 1493–1504. ACM, June...

    arXiv 2023

  3. [3]

    Easily accessible text-to- image generation amplifies demographic stereotypes at large scale

    Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to- image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493–1504, 2023

    work page 2023

  4. [4]

    FLUX.1-Schnell

    black-forest-labs. FLUX.1-Schnell. https://huggingface.co/black-forest-labs/ FLUX.1-schnell, 2024. Accessed: 2025-05-12

    work page 2024

  5. [5]

    Naruto blip captions

    Eole Cervenka. Naruto blip captions. https://huggingface.co/datasets/lambdalabs/ naruto-blip-captions/, 2022

    2022

  6. [6]

    Can editing llms inject harm?arXiv preprint arXiv:2407.20224, 2024

    Canyu Chen, Baixiang Huang, Zekun Li, Zhaorun Chen, Shiyang Lai, Xiongxiao Xu, Jia-Chen Gu, Jindong Gu, Huaxiu Yao, Chaowei Xiao, et al. Can editing llms inject harm?arXiv preprint arXiv:2407.20224, 2024

    arXiv 2024

  7. [7]

    Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models

    Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3043–3054, 2023

    2023

  8. [8]

    Openbias: Open-set bias detection in text-to- image generative models

    Moreno D’Incà, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vidit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. Openbias: Open-set bias detection in text-to- image generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12225–12235, 2024

    work page 2024

Show all 48 references
  1. [9]

    Pure: Turning polysemantic neurons into pure features by identifying relevant circuits

    Maximilian Dreyer, Erblina Purelku, Johanna Vielhaben, Wojciech Samek, and Sebastian La- puschkin. Pure: Turning polysemantic neurons into pure features by identifying relevant circuits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page...

    2024

  2. [10]

    Fleiss’ kappa statistic without paradoxes

    Rosa Falotico and Piero Quatto. Fleiss’ kappa statistic without paradoxes. Quality & Quantity, 49:463–470, 2015

    2015

  3. [11]

    Measuring nominal scale agreement among many raters

    Joseph L Fleiss. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378, 1971

    1971

  4. [12]

    Fair diffusion: Instructing text-to-image generation models on fairness, 2023

    Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness, 2023. URLhttps://arxiv.org/abs/2302.10893

    2023 arXiv

  5. [13]

    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 (ICCV), pages 2426–2436, October 2023

    2023

  6. [14]

    Unified concept editing in diffusion models

    Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzy´nska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111–5120, 2024. 11

    2024

  7. [15]

    Harm amplification in text-to-image models

    Susan Hao, Renee Shelby, Yuchi Liu, Hansa Srinivasan, Mukul Bhutani, Burcu Karagol Ayan, Ryan Poplin, Shivani Poddar, and Sarah Laszlo. Harm amplification in text-to-image models. arXiv preprint arXiv:2402.01787, 2024

    2024 arXiv

  8. [16]

    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. arXiv preprint arXiv:2104.08718, 2021

    2021 arXiv

  9. [17]

    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. Advances in neural information processing systems, 30, 2017

    2017

  10. [18]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1 (2):3, 2022

    2022

  11. [19]

    Harmful fine-tuning attacks and defenses for large language models: A survey

    Tiansheng Huang, Sihao Hu, Fatih Ilhan, Selim Furkan Tekin, and Ling Liu. Harmful fine-tuning attacks and defenses for large language models: A survey. arXiv preprint arXiv:2409.18169, 2024

    2024 arXiv

  12. [20]

    Ai alignment: A comprehensive survey

    Jiaming Ji, Tianyi Qiu, Boyuan Chen, Borong Zhang, Hantao Lou, Kaile Wang, Yawen Duan, Zhonghao He, Jiayi Zhou, Zhaowei Zhang, et al. Ai alignment: A comprehensive survey. arXiv preprint arXiv:2310.19852, 2023

    2023 arXiv

  13. [21]

    Rethinking training for de-biasing text-to-image generation: Unlocking the potential of stable diffusion, 2025

    Eunji Kim, Siwon Kim, Minjun Park, Rahim Entezari, and Sungroh Yoon. Rethinking training for de-biasing text-to-image generation: Unlocking the potential of stable diffusion, 2025. URL https://arxiv.org/abs/2408.12692

    2025 arXiv

  14. [22]

    Safety alignment backfires: Preventing the re-emergence of suppressed concepts in fine-tuned text-to-image diffusion models

    Sanghyun Kim, Moonseok Choi, Jinwoo Shin, and Juho Lee. Safety alignment backfires: Preventing the re-emergence of suppressed concepts in fine-tuned text-to-image diffusion models. arXiv preprint arXiv:2412.00357, 2024

    2024 arXiv

  15. [23]

    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

  16. [24]

    Self-discovering inter- pretable diffusion latent directions for responsible text-to-image generation

    Hang Li, Chengzhi Shen, Philip Torr, V olker Tresp, and Jindong Gu. Self-discovering inter- pretable diffusion latent directions for responsible text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024

    2024

  17. [25]

    BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors,Proceedings of...

    2022

  18. [26]

    Word-level explanations for analyzing bias in text-to-image models

    Alexander Lin, Lucas Monteiro Paes, Sree Harsha Tanneru, Suraj Srinivas, and Himabindu Lakkaraju. Word-level explanations for analyzing bias in text-to-image models. arXiv preprint arXiv:2306.05500, 2023

    2023 arXiv

  19. [27]

    Dora: Weight-decomposed low-rank adaptation

    Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. arXiv preprint arXiv:2402.09353, 2024

    2024 arXiv

  20. [28]

    Catastrophic interference in connectionist networks: The sequential learning problem

    Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, volume 24, pages 109–165. Elsevier, 1989

    1989

  21. [29]

    Locating and editing factual associations in GPT

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems, 35, 2022. 12

    2022

  22. [30]

    Mass editing memory in a transformer

    Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass editing memory in a transformer. The Eleventh International Conference on Learning Representations (ICLR), 2023

    2023

  23. [31]

    Prodigy: an expeditiously adaptive parameter- free learner

    Konstantin Mishchenko and Aaron Defazio. Prodigy: an expeditiously adaptive parameter- free learner. In Proceedings of the 41st International Conference on Machine Learning, pages 35779–35804, 2024

    2024

  24. [32]

    Social biases through the text-to-image generation lens

    Ranjita Naik and Besmira Nushi. Social biases through the text-to-image generation lens. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 786–808, 2023

    2023

  25. [33]

    Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks

    Anh Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks. arXiv preprint arXiv:1602.03616, 2016

    2016 arXiv

  26. [34]

    Disentangling neuron representations with concept vectors

    Laura O’Mahony, Vincent Andrearczyk, Henning Müller, and Mara Graziani. Disentangling neuron representations with concept vectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3770–3775, 2023

    2023

  27. [35]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7053–7061, 2023

    2023

  28. [36]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023

    2023 arXiv

  29. [38]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  30. [39]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  31. [40]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22522– 22531, June 2023

    2023

  32. [41]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522–22531, 2023

    2023

  33. [42]

    LAION-5b: An open large-scale dataset for training next gen- eration image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kacz- marczyk, and Jenia Jitsev....

    2022

  34. [43]

    Stablerep: Synthetic images from text-to-image models make strong visual representation learners

    Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. Advances in Neural Information Processing Systems, 36:48382–48402, 2023. 13

    2023

  35. [44]

    Survey of bias in text-to-image generation: Definition, evaluation, and mitigation

    Yixin Wan, Arjun Subramonian, Anaelia Ovalle, Zongyu Lin, Ashima Suvarna, Christina Chance, Hritik Bansal, Rebecca Pattichis, and Kai-Wei Chang. Survey of bias in text-to-image generation: Definition, evaluation, and mitigation. arXiv preprint arXiv:2404.01030, 2024

    2024 arXiv

  36. [45]

    Editing massive concepts in text-to-image diffusion models, 2024

    Tianwei Xiong, Yue Wu, Enze Xie, Yue Wu, Zhenguo Li, and Xihui Liu. Editing massive concepts in text-to-image diffusion models, 2024. URL https://arxiv.org/abs/2403. 13807

    2024

  37. [46]

    Position: Edit- ing large language models poses serious safety risks

    Paul Youssef, Zhixue Zhao, Daniel Braun, Jörg Schlötterer, and Christin Seifert. Position: Edit- ing large language models poses serious safety risks. In Forty-second International Conference on Machine Learning Position Paper Track, 2025. URL https://openreview.net/forum? id=...

    2025

  38. [47]

    A comprehensive study of knowledge editing for large language models

    Ningyu Zhang, Yunzhi Yao, Bozhong Tian, Peng Wang, Shumin Deng, Mengru Wang, Zekun Xi, Shengyu Mao, Jintian Zhang, Yuansheng Ni, et al. A comprehensive study of knowledge editing for large language models. arXiv preprint arXiv:2401.01286, 2024

    2024 arXiv

  39. [48]

    Gender bias in coreference resolution: Evaluation and debiasing methods

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu...

    2018

  40. [49]

    A CEO is giving a talk

    Jincheng Zhong, Xingzhuo Guo, Jiaxiang Dong, and Mingsheng Long. Diffusion tuning: Transferring diffusion models via chain of forgetting. arXiv preprint arXiv:2406.00773, 2024. 14 8 Limitations We validate our findings on three widely used text-to-image (T2I) diffusion models,...

    2024 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.