pith. machine review for the scientific record. sign in

arxiv: 2604.10032 · v1 · submitted 2026-04-11 · 💻 cs.LG · cs.AI

Recognition: 2 theorem links

· Lean Theorem

Closed-Form Concept Erasure via Double Projections

Chi Zhang , Jingpu Cheng , Zhixian Wang , Ping Liu
Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:40 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords concept erasuregenerative modelsdiffusion modelslinear transformationmodel editingconcept removalclosed-form methodStable Diffusion
0
0 comments X

The pith

A closed-form linear transformation using two sequential projections erases target concepts from generative models without training or optimization.

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

The paper introduces a deterministic method for removing unwanted concepts such as specific objects or styles from pretrained image generators. It does this by first applying a proxy projection that targets the unwanted direction in feature space and then performing a second transformation constrained to the left null space of known non-target directions. This produces an exact analytical update to the model that avoids the iterative optimization common in prior techniques. A reader would care because the procedure runs in seconds, is geometrically interpretable, and experiments indicate it removes the target at least as effectively as existing methods while distorting unrelated concepts less. The approach is presented as a lightweight, drop-in tool for safer model editing across diffusion and flow-matching architectures.

Core claim

The authors establish that concept erasure reduces to a pair of closed-form linear operations: first compute a proxy projection of the target concept, then apply a transformation lying in the left null space of the directions that represent concepts to be preserved. The resulting weight update removes the target concept from the model's output distribution while leaving the geometry of non-target directions intact, all without any gradient-based fine-tuning or iterative search.

What carries the argument

The double-projection operator: an initial proxy projection onto the target concept direction followed by a constrained linear map inside the left null space of known non-target directions.

If this is right

  • Erasure becomes a one-time, deterministic calculation that finishes in seconds rather than requiring hours of optimization.
  • Non-target concepts remain more faithfully represented than under state-of-the-art iterative erasure methods.
  • The same procedure works on multiple Stable Diffusion variants and on flow-matching models such as FLUX.
  • The update is fully analytical and therefore reproducible across runs and hardware.

Where Pith is reading between the lines

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

  • If linear directions prove sufficient for many concepts, the same null-space construction could support rapid removal of additional model behaviors beyond image content.
  • The speed of the method opens the possibility of on-the-fly editing inside interactive creative applications.
  • Extending the approach to non-image modalities would test whether the linear-representability premise generalizes beyond diffusion and flow-matching image models.

Load-bearing premise

Target concepts exist as distinct linear directions in the model's feature space that can be isolated and subtracted without altering the representation of unrelated concepts.

What would settle it

After applying the double-projection update, generate images from prompts that explicitly request the erased concept and check whether those images continue to contain the concept at rates comparable to the original model.

Figures

Figures reproduced from arXiv: 2604.10032 by Chi Zhang, Jingpu Cheng, Ping Liu, Zhixian Wang.

Figure 1
Figure 1. Figure 1: Visualization of concept erasure on “Cassette player”. Results indicated that the target concept (column 1) is effectively sup [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Computation time comparison across erasure methods [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Concept erasure on “Cassette Player” with anchor concept “Box”. The first column shows the target concept to be erased. [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Concept erasure on a few target concepts with FLUX. Results demonstrate that the proposed DP method successfully suppresses [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗
read the original abstract

While modern generative models such as diffusion-based architectures have enabled impressive creative capabilities, they also raise important safety and ethical risks. These concerns have led to growing interest in concept erasure, the process of removing unwanted concepts from model representations. Existing approaches often achieve strong erasure performance but rely on iterative optimization and may inadvertently distort unrelated concepts. In this work, we present a simple yet principled alternative: a linear transformation framework that achieves concept erasure analytically, without any training. Our method adapts a pretrained model through two sequential, closed-form steps: first, computing a proxy projection of the target concept, and second, applying a constrained transformation within the left null space of known concept directions. This design yields a deterministic and geometrically interpretable procedure for safe, efficient, and theory-grounded concept removal. Across a wide range of experiments, including object and style erasure on multiple Stable Diffusion variants and the flow-matching model (FLUX), our approach matches or surpasses the performance of state-of-the-art methods while preserving non-target concepts more faithfully. Requiring only a few seconds to apply, it offers a lightweight and drop-in tool for controlled model editing, advancing the goal of safer and more responsible generative models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes a closed-form, training-free linear transformation for concept erasure in generative models. It computes a proxy projection onto a target concept direction followed by a second linear map constrained to the left nullspace of known directions, claiming this erases the target while leaving non-target concepts unchanged. Experiments on object and style erasure across Stable Diffusion variants and FLUX report performance matching or exceeding state-of-the-art methods with improved fidelity on unrelated concepts.

Significance. If the algebraic construction provides a parameter-free guarantee of erasure without distortion, the method would be a lightweight, deterministic tool for model editing that avoids the computational cost and potential side-effects of optimization-based approaches. The emphasis on geometric interpretability and applicability to both diffusion and flow-matching architectures strengthens its potential utility for safety-critical editing tasks.

major comments (3)
  1. [§3] §3 (Method), the double-projection construction: the claim that the second map is identity on non-target subspaces requires that the known directions form a basis whose left nullspace contains an isometry on the orthogonal complement. No proof or bound is given showing that residual target-concept norm is zero when the proxy direction only approximately spans the target (as is typical in entangled UNet/FLUX features).
  2. [§4] §4 (Experiments), quantitative tables: the reported parity or gains over SOTA are stated without accompanying residual-concept-norm measurements or distortion metrics on held-out non-target concepts. Without these, the assertion that non-target concepts are preserved 'more faithfully' cannot be evaluated against the geometric guarantee.
  3. [§3.2] §3.2 (Proxy projection definition): the procedure is described as 'parameter-free,' yet the choice of which directions are treated as 'known' and how the proxy direction is extracted from data implicitly introduces modeling choices whose sensitivity is not analyzed.
minor comments (2)
  1. [§3] Notation for the left-nullspace projection operator is introduced without an explicit matrix formula or pseudocode; adding one would improve reproducibility.
  2. [Abstract and §4] The abstract and introduction cite 'multiple Stable Diffusion variants' but the experimental section should list the exact model checkpoints and layer indices used for feature extraction.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the thoughtful and detailed comments on our manuscript. We address each major comment point by point below, indicating planned revisions where appropriate.

read point-by-point responses

  1. Referee: [§3] §3 (Method), the double-projection construction: the claim that the second map is identity on non-target subspaces requires that the known directions form a basis whose left nullspace contains an isometry on the orthogonal complement. No proof or bound is given showing that residual target-concept norm is zero when the proxy direction only approximately spans the target (as is typical in entangled UNet/FLUX features).

    Authors: We appreciate this observation regarding the theoretical guarantee. The double-projection construction ensures the second map acts as the identity on the orthogonal complement to the known directions precisely when the proxy direction spans the target concept. For the common case of approximate proxies arising from entangled features, the manuscript provides no formal bound on residual target-concept norm. We will revise §3 to explicitly state the exact conditions under which the identity property holds and augment the discussion with empirical residual-norm measurements from our experiments to characterize behavior in the approximate setting. revision: partial

  2. Referee: [§4] §4 (Experiments), quantitative tables: the reported parity or gains over SOTA are stated without accompanying residual-concept-norm measurements or distortion metrics on held-out non-target concepts. Without these, the assertion that non-target concepts are preserved 'more faithfully' cannot be evaluated against the geometric guarantee.

    Authors: We agree that the current quantitative tables would be strengthened by these additional metrics. We will update §4 to include residual target-concept norm after erasure as well as distortion metrics (e.g., concept similarity on held-out non-target concepts) across the Stable Diffusion and FLUX experiments, allowing direct comparison against the geometric claims. revision: yes

  3. Referee: [§3.2] §3.2 (Proxy projection definition): the procedure is described as 'parameter-free,' yet the choice of which directions are treated as 'known' and how the proxy direction is extracted from data implicitly introduces modeling choices whose sensitivity is not analyzed.

    Authors: The term 'parameter-free' in the manuscript refers specifically to the closed-form, training-free nature of the linear transformation once directions are selected. We acknowledge that identifying known directions and computing the proxy from data constitute modeling choices. We will revise §3.2 to clarify this distinction and discuss the sensitivity of results to these choices, drawing on the robustness observed across our reported experiments. revision: partial

standing simulated objections not resolved
  • A formal proof or bound on residual target-concept norm when the proxy direction only approximately spans the target concept.

Circularity Check

0 steps flagged

No circularity: derivation is algebraic and self-contained

full rationale

The paper's central construction is a two-step linear map (proxy projection onto a target direction followed by a transformation constrained to the left nullspace of known directions) that is derived directly from the definitions of projection and nullspace operators. No parameter is fitted to data and then relabeled as a prediction, no result is defined in terms of itself, and no load-bearing premise rests on a self-citation chain. The algebraic steps are presented as closed-form and deterministic; any empirical success or failure is therefore an external test of the modeling assumptions rather than a tautology internal to the derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on standard linear-algebra assumptions about concept directions in feature space and the ability to compute exact projections from pretrained weights; no free parameters, invented entities, or ad-hoc axioms are introduced in the abstract description.

axioms (2)
  • domain assumption Target concepts admit linear directional representations in the model's activation space.
    Required for the proxy projection and null-space steps to function as described.
  • domain assumption The left null space of known concept directions can be used to constrain the transformation without affecting non-target behavior.
    Central to the claim of faithful preservation of unrelated concepts.

pith-pipeline@v0.9.0 · 5505 in / 1281 out tokens · 52999 ms · 2026-05-10T16:40:48.013113+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

76 extracted references · 21 canonical work pages · 5 internal anchors

  1. [1]

    Albergo and Eric Vanden-Eijnden

    Michael S. Albergo and Eric Vanden-Eijnden. Building nor- malizing flows with stochastic interpolants. InICLR, 2023. 2

    2023

  2. [2]

    Sarah Andersen, Kelly McKernan, and Karla Ortiz et al

    Sarah Andersen, Kelly McKernan, and Karla Ortiz. Sarah Andersen, Kelly McKernan, and Karla Ortiz et al. v. Stability AI Ltd. et al., 2023. Case No. 3:2023cv00201, US District Court for the Northern District of California. 2

    2023

  3. [3]

    Stephen Batifol, Andreas Blattmann, Frederic Boesel, Sak- sham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, et al. Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv e-prints, pages arXiv–2506,

  4. [4]

    Bezdek and Richard J

    James C. Bezdek and Richard J. Hathaway. Convergence of alternating optimization.Neural, Parallel & Scientific Com- putations, 11(4):351–368, 2003. 4

    2003

  5. [5]

    Easily acces- sible text-to-image generation amplifies demographic stereo- types 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 acces- sible text-to-image generation amplifies demographic stereo- types at large scale. InProceedings of the 2023 ACM Confer- ence on Fairness, Accountability, and Transparency (FAccT), pages...

    2023

  6. [6]

    Bovik.Handbook of Image and Video Processing

    Alan C. Bovik.Handbook of Image and Video Processing. Academic Press, 2000. 8, 19

    2000

  7. [7]

    Distributed optimization and sta- tistical learning via the alternating direction method of mul- tipliers.Foundations and Trends® in Machine learning, 3 (1):1–122, 2011

    Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed optimization and sta- tistical learning via the alternating direction method of mul- tipliers.Foundations and Trends® in Machine learning, 3 (1):1–122, 2011. 4

    2011

  8. [8]

    Fantastic tar- gets for concept erasure in diffusion models and where to find them

    Anh Bui, Trang Vu, Long Vuong, Trung Le, Paul Montague, Tamas Abraham, Junae Kim, and Dinh Phung. Fantastic tar- gets for concept erasure in diffusion models and where to find them. 2025. 1, 2, 5, 6, 16

    2025

  9. [9]

    Extracting training data from large lan- guage models

    Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-V oss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, Alina Oprea, and Jacob Steinhardt. Extracting training data from large lan- guage models. InProceedings of the 30th USENIX Security Symposium, 2021. 1

    2021

  10. [10]

    Quanti- fying memorization across neural language models

    Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quanti- fying memorization across neural language models. InICLR,

  11. [11]

    Extracting training data from diffu- sion models

    Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagiel- ski, Vikash Sehwag, Florian Tram`er, Borja Balle, Daphne Ip- polito, and Eric Wallace. Extracting training data from diffu- sion models. In32nd USENIX Security Symposium (USENIX Security), 2023. 2

    2023

  12. [12]

    Trace: Trajectory-constrained concept erasure in diffusion models.arXiv preprint arXiv:2505.23312, 2025

    Finn Carter. Trace: Trajectory-constrained concept erasure in diffusion models.arXiv preprint arXiv:2505.23312, 2025. 2

    work page arXiv 2025

  13. [13]

    Con- ceptprune: Concept editing in diffusion models via skilled neuron pruning

    Ruchika Chavhan, Da Li, and Timothy Hospedales. Con- ceptprune: Concept editing in diffusion models via skilled neuron pruning. InICLR, 2025. 1, 2, 6

    2025

  14. [14]

    Ma- chine unlearning under retain–forget entanglement

    Jingpu Cheng, Ping Liu, Qianxiao Li, and Chi Zhang. Ma- chine unlearning under retain–forget entanglement. InICLR,

  15. [15]

    Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts.arXiv preprint arXiv:2309.06135, 2023

    Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, Pin- Yu Chen, and Wei-Chen Chiu. Prompting4debugging: Red- teaming text-to-image diffusion models by finding problem- atic prompts.arXiv preprint arXiv:2309.06135, 2023. 2

    work page arXiv 2023

  16. [16]

    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. InICCV, 2023. 2

    2023

  17. [17]

    SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders

    Bartosz Cywi ´nski and Kamil Deja. SAeUron: Interpretable concept unlearning in diffusion models with sparse autoen- coders.arXiv preprint arXiv:2501.18052, 2025. 2

    work page arXiv 2025

  18. [18]

    Lo- RAShop: Training-free multi-concept image generation and editing with rectified flow transformers.arXiv preprint arXiv:2505.23758, 2025

    Yusuf Dalva, Hidir Yesiltepe, and Pinar Yanardag. Lo- RAShop: Training-free multi-concept image generation and editing with rectified flow transformers.arXiv preprint arXiv:2505.23758, 2025. 2

    work page arXiv 2025

  19. [19]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 1, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  20. [20]

    Al- phaedit: Null-space constrained model editing for language models

    Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Jie Shi, Xiang Wang, Xiangnan He, and Tat-Seng Chua. Al- phaedit: Null-space constrained model editing for language models. InICLR, 2025. 2

    2025

  21. [21]

    John Wiley & Sons, 1999

    Gerald B Folland.Real analysis: modern techniques and their applications. John Wiley & Sons, 1999. 3

    1999

  22. [22]

    An image is worth one word: Personalizing text-to-image gen- eration using textual inversion

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image gen- eration using textual inversion. InICLR, 2022. 2

    2022

  23. [23]

    Erasing concepts from diffusion models

    Rohit Gandikota, Joanna Materzynska, Jaden Fiotto- Kaufman, and David Bau. Erasing concepts from diffusion models. InICCV, 2023. 1, 2, 6

    2023

  24. [24]

    Unified concept editing in dif- fusion models

    Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzy´nska, and David Bau. Unified concept editing in dif- fusion models. InWACV, 2024. 1, 2, 3, 4, 5, 6, 7

    2024

  25. [25]

    Eraseanything: Enabling concept erasure in rectified flow transformers

    Daiheng Gao, Shilin Lu, Wenbo Zhou, Jiaming Chu, Jie Zhang, Mengxi Jia, Bang Zhang, Zhaoxin Fan, and Weiming Zhang. Eraseanything: Enabling concept erasure in rectified flow transformers. InICML, 2025. 2

    2025

  26. [26]

    Dumo: Dual encoder modulation network for precise concept erasure

    Feng Han, Kai Chen, Chao Gong, Zhipeng Wei, Jingjing Chen, and Yu-Gang Jiang. Dumo: Dual encoder modulation network for precise concept erasure. InAAAI, 2025. 2

    2025

  27. [27]

    Vce: Safe autoregressive image gen- eration via visual contrast exploitation.arXiv preprint arXiv:2509.16986, 2025

    Feng Han, Chao Gong, Zhipeng Wei, Jingjing Chen, and Yu-Gang Jiang. Vce: Safe autoregressive image gen- eration via visual contrast exploitation.arXiv preprint arXiv:2509.16986, 2025. 2

    work page arXiv 2025

  28. [28]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InCVPR,

  29. [29]

    A single neuron works: Precise concept erasure in text-to-image diffusion models.arXiv preprint arXiv:2509.21008, 2025

    Qinqin He, Jiaqi Weng, Jialing Tao, and Hui Xue. A sin- gle neuron works: Precise concept erasure in text-to-image diffusion models.arXiv preprint arXiv:2509.21008, 2025. 2

    work page arXiv 2025

  30. [30]

    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. In NeurIPS, 2023. 2

    2023

  31. [31]

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

    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 equilib- rium. InNeurIPS, 2017. 3, 8, 19

    2017

  32. [32]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. InNeurIPS, 2020. 1, 2, 3

    2020

  33. [33]

    Lora: Low-rank adaptation of large language models

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

    2022

  34. [34]

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

    Chi-Pin Huang, Kai-Po Chang, Chung-Ting Tsai, Yung- Hsuan Lai, Fu-En Yang, and Yu-Chiang Frank Wang. Re- celer: Reliable concept erasing of text-to-image diffusion models via lightweight erasers. InECCV, 2024. 1, 2

    2024

  35. [35]

    AI porn is easy to make now

    Tatum Hunter. AI porn is easy to make now. For women, that’s a nightmare. The Washington Post, 2023. 2

    2023

  36. [36]

    Concept steerers: Leveragingk-sparse autoencoders for controllable genera- tions.arXiv preprint arXiv:2501.19066, 2025

    Dahye Kim and Deepti Ghadiyaram. Concept steerers: Leveragingk-sparse autoencoders for controllable genera- tions.arXiv preprint arXiv:2501.19066, 2025. 2

    work page arXiv 2025

  37. [37]

    Towards safe self- distillation of internet-scale text-to-image diffusion models

    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. 2

    work page arXiv 2023

  38. [38]

    Ablating con- cepts in text-to-image diffusion models

    Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating con- cepts in text-to-image diffusion models. InICCV, 2023. 2

    2023

  39. [39]

    Multi-concept customization of text-to-image diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. InCVPR, 2023. 2

    2023

  40. [40]

    Local- ized concept erasure for text-to-image diffusion models us- ing training-free gated low-rank adaptation

    Byung Hyun Lee, Sungjin Lim, and Se Young Chun. Local- ized concept erasure for text-to-image diffusion models us- ing training-free gated low-rank adaptation. InCVPR, 2025. 2

    2025

  41. [41]

    Algorithms for non- negative matrix factorization

    Daniel Lee and H Sebastian Seung. Algorithms for non- negative matrix factorization. InNeurIPS, 2000. 4

    2000

  42. [42]

    Flow Matching for Generative Modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling.arXiv preprint arXiv:2210.02747, 2022. 1, 2, 6, 7

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  43. [43]

    Erased or dormant? rethink- ing concept erasure through reversibility.arXiv preprint arXiv:2505.16174, 2025

    Ping Liu and Chi Zhang. Erased or dormant? rethink- ing concept erasure through reversibility.arXiv preprint arXiv:2505.16174, 2025. 2

    work page arXiv 2025

  44. [44]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InICLR, 2023. 2

    2023

  45. [45]

    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. InCVPR, 2024. 1, 2

    2024

  46. [46]

    arXiv preprint arXiv:2303.11408 (2023)

    Alexandra Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Analyzing so- cietal representations in diffusion models.arXiv preprint arXiv:2303.11408, 2023. 2

    work page arXiv 2023

  47. [47]

    On the reciprocal of the general algebraic matrix.Bulletin of the american mathematical society, 26: 294–295, 1920

    Eliakim H Moore. On the reciprocal of the general algebraic matrix.Bulletin of the american mathematical society, 26: 294–295, 1920. 4

    1920

  48. [48]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InICCV, 2023. 1

    2023

  49. [49]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InICML, 2021. 3, 6

    2021

  50. [50]

    Hierarchical Text-Conditional Image Generation with CLIP Latents

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

    work page internal anchor Pith review arXiv

  51. [51]

    Red-teaming the stable diffusion safety filter.arXiv preprint arXiv:2210.04610, 2022

    Javier Rando, Daniel Paleka, David Lindner, Lennard Heim, and Florian Tram`er. Red-teaming the Stable Diffusion safety filter.arXiv preprint arXiv:2210.04610, 2022. 2

    work page arXiv 2022

  52. [52]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InCVPR, 2022. 1, 2

    2022

  53. [53]

    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. InCVPR, 2023. 2

    2023

  54. [54]

    Raising the cost of malicious AI-powered image editing.arXiv preprint arXiv:2302.06588, 2023

    Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, An- drew Ilyas, and Aleksander Madry. Raising the cost of malicious AI-powered image editing.arXiv preprint arXiv:2302.06588, 2023. 2

    work page arXiv 2023

  55. [55]

    Safe latent diffusion: Mitigating inappro- priate degeneration in diffusion models

    Patrick Schramowski, Manuel Brack, Bj ¨orn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappro- priate degeneration in diffusion models. InCVPR, 2023. 2

    2023

  56. [56]

    Glaze: Protecting artists from style mimicry by text-to-image models.arXiv preprint arXiv:2302.04222, 2023

    Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models.arXiv preprint arXiv:2302.04222, 2023. 2

    work page arXiv 2023

  57. [57]

    Instant- booth: Personalized text-to-image generation without test- time finetuning

    Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instant- booth: Personalized text-to-image generation without test- time finetuning. InCVPR, 2024. 2

    2024

  58. [58]

    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. InCVPR,

  59. [59]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 1

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  60. [60]

    Rectifid: Personalizing rectified flow with an- chored classifier guidance

    Zhicheng Sun, Zhenhao Yang, Yang Jin, Haozhe Chi, Kun Xu, Liwei Chen, Hao Jiang, Yang Song, Kun Gai, and Yadong Mu. Rectifid: Personalizing rectified flow with an- chored classifier guidance. InNeurIPS, 2024. 2

    2024

  61. [61]

    Unified diffusion- based rigid and non-rigid editing with text and image guid- ance

    Jiacheng Wang, Ping Liu, and Wei Xu. Unified diffusion- based rigid and non-rigid editing with text and image guid- ance. InICME, 2024. 2

    2024

  62. [62]

    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 Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004. 8, 19

    2004

  63. [63]

    Erasing concepts, steering generations: A comprehensive survey of concept suppression.arXiv preprint arXiv:2505.19398, 2025

    Yiwei Xie, Ping Liu, and Zheng Zhang. Erasing concepts, steering generations: A comprehensive survey of concept suppression.arXiv preprint arXiv:2505.19398, 2025. 2

    work page arXiv 2025

  64. [64]

    Video- eraser: Concept erasure in text-to-video diffusion models

    Naen Xu, Jinghuai Zhang, Changjiang Li, Zhi Chen, Chunyi Zhou, Qingming Li, Tianyu Du, and Shouling Ji. Video- eraser: Concept erasure in text-to-video diffusion models. In EMNLP, 2025. 2

    2025

  65. [65]

    Pruning for robust concept erasing in diffusion models

    Tianyun Yang, Juan Cao, and Chang Xu. Pruning for robust concept erasing in diffusion models. InNeurIPS Workshops,

  66. [66]

    T2vunlearning: A concept erasing method for text-to-video diffusion models.arXiv preprint arXiv:2505.17550, 2025

    Xiaoyu Ye, Songjie Cheng, Yongtao Wang, Yajiao Xiong, and Yishen Li. T2vunlearning: A concept erasing method for text-to-video diffusion models.arXiv preprint arXiv:2505.17550, 2025. 2

    work page arXiv 2025

  67. [67]

    Scaling Autoregressive Models for Content-Rich Text-to-Image Generation

    Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gun- jan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin- fei Yang, Burcu Karagol Ayan, et al. Scaling autoregres- sive models for content-rich text-to-image generation.arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 1

    work page internal anchor Pith review arXiv 2022

  68. [68]

    Parameter-efficient fine-tuning with controls

    Chi Zhang, Cheng Jingpu, Yanyu Xu, and Qianxiao Li. Parameter-efficient fine-tuning with controls. InForty-first International Conference on Machine Learning, 2024. 2

    2024

  69. [69]

    From weight-based to state-based fine-tuning: Further mem- ory reduction on lora with parallel control

    Chi Zhang, REN Lianhai, Jingpu Cheng, and Qianxiao Li. From weight-based to state-based fine-tuning: Further mem- ory reduction on lora with parallel control. InForty-second International Conference on Machine Learning, 2025. 2

    2025

  70. [70]

    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. InCVPR, pages 1755–1764,

  71. [71]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InCVPR (CVPR), 2018. 8, 19

    2018

  72. [72]

    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. InNeurIPS,

  73. [73]

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

    Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yi- hua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To gener- ate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. InECCV, 2024. 2

    2024

  74. [74]

    cassette player,

    Mengnan Zhao, Lihe Zhang, Tianhang Zheng, Yuqiu Kong, and Baocai Yin. Separable multi-concept erasure from dif- fusion models.arXiv preprint arXiv:2402.05947, 2024. 2 Appendix Catalogue A . Proof for Theorems 13 A.1 . Additional Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 A.2 . Proof for The...

    work page arXiv 2024

  75. [75]

    tower” and “factory

    A multi-vector safe region constructed from the concepts “tower” and “factory”

  76. [76]

    Cassette Player

    A single-vector safe region using only “factory” as the anchor (as in the main experiments). Table 7. Ablation study on the construction of the safe subspace for the target conceptChurch. We compare a multi-vector safe region (Tower + Factory) with a single-vector anchor (Factory). Left block reports erasure performance; right block reports preservation q...