UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning

Maciej Zieba; Maksym Petrenko; Piotr W\'ojcik; Przemys{\l}aw Spurek; Wojciech Gromski

REVIEW 2 major objections 5 minor 16 references

UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning

T0 review · 2 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read UnHype claims that one hypernetwork, given a concept's text embedding, can generate LoRA weights on the fly that erase that concept — and even unseen concepts — without per-concept fine-tuning.

desk verdict New combination of hypernetwork gradient-matching and LoRA unlearning, but the zero-shot synonym claim is undercut by train/test overlap in the object-erasure evaluation. read the letter →

arxiv 2602.03410 v2 pith:DY3AVV7V submitted 2026-02-03 cs.CV

Piotr Wójcik , Maksym Petrenko , Wojciech Gromski , Przemys{l}aw Spurek , Maciej Zieba This is my paper · ORCID

classification cs.CV

keywords machineunlearningdiffusionmodelshypernetworkslow-rankadaptationconcepterasureCLIPembeddingszero-shotgeneralizationtext-to-imagegeneration

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 sets out to turn machine unlearning in text-to-image models from a per-concept fine-tuning chore into an amortized, one-network operation. It presents UnHype, in which a hypernetwork takes a concept's text embedding and a step index and outputs LoRA weight updates for the diffusion model. Training matches the hypernetwork's own trajectory through weight space to single stochastic-gradient steps of an unlearning loss, so no pre-computed LoRA weights are needed. At inference, a forbidden prompt makes the network emit weights that steer generation away from that concept, while unrelated prompts make it emit near-zero weights. If this works, erasing objects, celebrities, or explicit content becomes cheaper and scalable, with zero-shot generalization to synonyms and unseen concepts.

What carries the argument

The load-bearing object is an MLP hypernetwork that, given a concept's text embedding and an unlearning step index, outputs the full set of LoRA weights for the model's conditioning layers. Training uses the gradient-matching removal loss: the hypernetwork's own step from s to s+1 must equal one SGD step of the unlearning task loss, which steers the adapted model away from the target concept and toward a replacement concept. A retention loss anchors non-target concepts to zero weights, making the hypernetwork behave as a semantic switch that leaves unrelated generations untouched. This machinery converts per-concept fine-tuning into amortized, zero-shot adapter generation.

What would settle it

Train UnHype on, say, five object concepts, then prompt the final model with a sixth concept (e.g., 'zebra') that never appeared in training and count how often recognisable zebra images are still produced; if the rate is close to the no-unlearning baseline, zero-shot generalisation fails. A second check is to fix the training set and increase the number of simultaneously erased concepts: if image quality on retained content degrades sharply or erasure efficacy falls, the amortisation claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that unlearning can be modelled as a continuous, semantically-conditioned process rather than a static fine-tuning result. A hypernetwork maps a text embedding c and an unlearning step s to LoRA parameters θ_s. It is trained with a removal loss that forces the difference between consecutive outputs, θ_{s+1} − θ_s, to match the gradient of an unlearning task loss evaluated at the current weights, together with a retention loss that forces near-zero outputs for concepts that should be kept. At inference, one forward pass at the final step S produces the LoRA update; in a latent diffusion model the update is applied only to the conditional guidance pass, while in a flow-bas

Load-bearing premise

The load-bearing premise is that forcing the hypernetwork's successive outputs to match one small gradient step of the unlearning loss, at sampled points along its own path, is enough to teach one network to produce a valid unlearning path for any concept, including ones it has never seen, even though the replacement concept is not supplied at inference.

Editorial extensions

If this is right

A model owner can erase a new concept at inference time by feeding its text embedding to the hypernetwork, with no retraining and no stored per-concept adapter.
Many concepts can be erased by one hypernetwork: the paper trains a single model to suppress 100 celebrities at once while keeping 100 non-target identities intact.
The same framework transfers across architecture families, since applying the generated LoRA to the conditioning pass in a latent diffusion model vs. directly in a flow-based model is the only change.
Because the hypernetwork is continuous in embedding space, erasure can generalise to synonyms and paraphrases that never appeared in training, addressing a known weakness of prompt-specific unlearning.
Erasure strength becomes a matter of choosing the trajectory endpoint S, so one trained network can serve as a tunable, switchable unlearning module.

Reading between the lines

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

A natural stress test the paper does not run is to train on one cluster of concepts and hold out a semantically distant concept; if the zero-shot LoRA fails there, the 'unseen concepts' claim is really only about interpolating near training concepts, not open-set erasure.
Exposing the step index s at inference could let a user dial the strength of erasure continuously, turning the trajectory into a control knob that the paper only uses at its endpoint.
Since the task loss specifies a replacement concept, the same hypernetwork could be repurposed to redirect a concept toward different targets by swapping the mapping prompt, offering a cheap way to re-route content rather than merely delete it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Hi [Name],

You should know two things about UnHype. First, the core idea is a real novelty: a CLIP-conditioned hypernetwork trained with Hypernet Fields' gradient-matching objective to generate LoRA erasure weights on the fly, instead of per-concept fine-tuning. The training loss in Eq. (5)-(6) is coherent, and the stable training behavior on both SD and Flux is plausible. Second, the headline zero-shot claim is not supported by the experiments. The object-erasure 'generality' metric in Table 1 is measured on synonyms ('aircraft', 'plane', 'jet') that are verbatim training targets in Appendix A. So Table 1 only shows erasure of concepts the hypernetwork has seen; it says nothing about generalization to unseen embeddings. This is a load-bearing flaw because zero-shot adaptation is the paper's main differentiator.

What's good: the method is simple, parameter-efficient, and the multi-concept celebrity erasure (100 at once) is genuinely useful. The Flux results look strong, and the comparison to SAeUron's training time is a fair practical point. I also appreciate that the paper explicitly identifies its lineage (Hypernet Fields, UnGuide) rather than hiding it.

Soft spots, in proportion: the synonym-overlap issue is the big one; it needs a re-run with held-out synonyms. The loss weights lambda and gamma are never reported, and three Flux baselines are uncited. There are no error bars or sample sizes, which matters for a method with this much variance. The assumption that a single MLP can amortize full unlearning trajectories across the whole CLIP embedding space is unverified, but that's a modeling risk, not an error.

Who this is for: anyone working on machine unlearning or diffusion safety. The method deserves referee time — the idea is fresh, the math is sound, and the flaw is fixable. I would not desk reject, but I would send to reviewers with a clear request to rerun object erasure with held-out synonyms and to report the missing hyperparameters. Once that's done, this could be a solid paper.

My recommendation: engage with it, but don't cite the zero-shot claim yet.

Best, [Your name]

Referee Report

2 major / 5 minor

Summary. The paper introduces UnHype, a framework that trains a hypernetwork to generate LoRA weights for concept erasure in text-to-image diffusion models. The hypernetwork is conditioned on a CLIP text embedding and a continuous optimization step, and is trained by matching its trajectory in weight space to single-step SGD gradients of an UnGuide-style task loss, combined with a retention loss that forces near-zero weights for non-target concepts. At inference, the hypernetwork outputs final LoRA weights in one forward pass. The authors evaluate on Stable Diffusion 1.4 and Flux for object erasure, nudity erasure, and 100-celebrity removal, comparing against multiple baselines. The paper claims zero-shot generalization to unseen synonyms and scalable multi-concept erasure.

Significance. If the central claims held, the work would be a meaningful advance: it replaces per-concept LoRA fine-tuning with an amortized generator, supports simultaneous erasure of many concepts, and transfers across two diffusion architectures. The training objective is a coherent application of Hypernet Fields and avoids the need for precomputed LoRA targets; the authors also provide code and extensive benchmark comparisons. The main weakness is that the headline capability—generalization to concepts not seen during training—is not actually tested. Once that is established or the claims are appropriately weakened, the amortized training scheme itself remains a useful contribution.

major comments (2)

[§5.2, Appendix C.1 vs Appendix A] The Generality metric Acc_g is presented as evidence of 'unseen synonym' erasure (Table 1). However, the synonym prompts used in C.1 are verbatim training targets in Appendix A: for airplane, training includes 'a jet', 'a plane', 'an aircraft' and C.1 evaluates 'aircraft', 'plane', 'jet'; the same overlap holds for bird ('owl', 'warbler', 'songbird') and ship ('frigate', 'ocean liner', 'ferry'). Appendix A's statement that the training terms 'do not appear in the test set' is contradicted by the protocol. Consequently Table 1's Acc_g and H_o only demonstrate erasure of training concepts; the zero-shot generalization claim—restated in the abstract, introduction, and conclusion—is not empirically supported. Please re-run with held-out synonyms or remove the zero-shot claim.
[§4.2, Eqs. (4)–(5)] The task loss L_task is defined with a mapping concept c_m (e.g., airplane→ground), but the hypernetwork input is only the text embedding c (plus step s). The gradient-matching removal loss aligns the predicted trajectory with ∇_{θ_s} L_task for the fixed (c, c_m) pairs used in training. No analysis or experiment addresses whether this teaches H_φ to infer an appropriate mapping for truly unseen c at inference; the only test of generality (Appendix C.1) uses prompts already present in training. Thus the 'context-aware' and 'zero-shot' behaviour is an unverified capacity assumption, and the amortized mechanism's core advantage remains unmeasured.

minor comments (5)

[Table 2] The UnHype row in the nudity-erasure table is misaligned/garbled in the text ('UnHype 26032 01 41814.20 30.99'); please ensure the eight anatomical counts and totals are reported legibly.
[Table 4] FID/CLIP are reported as '-' for the standard UnHype variant; provide these numbers or justify their omission if they are not available.
[§5.4] For celebrity removal, the hypernetwork is conditioned on NV-Embed-v2 rather than the CLIP embedding. This conflicts with the paper's 'CLIP-guided' framing. Clarify whether the method is agnostic to the text encoder and whether this affects the comparisons in Table 4.
[Appendix A] The sentence 'For category erasure, we selected semantically related terms that do not appear in the test set' is incorrect given that these terms are used in the Generality evaluation (C.1). Correct the statement to acknowledge the overlap or revise the protocol.
[§5.2] The claim of 'superior generalization to unseen synonyms' is based on only three synonyms per class. Since these are shown to be training terms, either re-evaluate with genuinely held-out paraphrases or soften the claim.

Circularity Check

1 steps flagged · score 6.0 of 10

Zero-shot synonym erasure is evaluated on the exact training prompts; the central 'generalization' claim reduces to in-sample fitting.

fitted input called prediction [Appendix A (training prompts, Airplane Erasure) vs §5.2 / Appendix C.1 (Generality metric)]
"For category erasure, we selected semantically related terms that do not appear in the test set to ensure effective synonym removal. [Appendix A] ... Airplane Erasure Target concepts: an airplane, a jet, a plane, an aircraft. [C.1] Generality (Acc_g): ... preparing three synonyms for each object class (e.g., 'aircraft', 'plane', and 'jet' for airplane)."

The only measure intended to show zero-shot synonym generalization (Acc_g) uses words that appear verbatim among the training targets in Appendix A ('jet', 'plane', 'aircraft' for airplane), directly contradicting the asserted train/test split. Table 1's Acc_e/Acc_g therefore demonstrate erasure of seen training prompts, not generalization to unseen CLIP embeddings. Since §4.3 and the Conclusion rest the method's contribution on 'zero-shot generalization to unseen synonyms', this prediction is in-sample by construction; the claimed out-of-distribution result is unevidenced.

full rationale

The core training objective (Eq. 4-6) is not circular in an equation-level sense: Lremove matches the hypernetwork's trajectory gradient to the task-loss gradient evaluated at its own outputs, with the target ϵtarget defined from the frozen base model; no benchmark labels are used as supervision, and the hypernetwork is not fitted to the reported CLIP/NudeNet/GCD metrics. The same-group UnGuide citation supplies the task loss but is not invoked as evidence for the final result, so it does not load-bearing circularity. However, the paper's central differentiator—zero-shot, context-aware erasure of unseen concepts—is only tested with the Generality metric, and those synonym prompts are identical to Appendix A training targets. Thus the headline empirical prediction reduces to an in-sample fit. This is an evaluation/validation circularity rather than a derivation-level one; the underlying gradient-matching mechanism may be sound but the paper's main claim is currently unsupported.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The framework rests on two borrowed training principles (Hypernet Fields gradient matching and UnGuide's guided-regression task loss) plus several unreported hyperparameters. The only new entity is the CLIP-conditioned hypernetwork; there is no independent evidence for its zero-shot generalization because tested synonyms were used in training.

free parameters (6)

S (trajectory endpoint) = 300 for all experiments
Number of unlearning trajectory steps used at inference; no systematic study is presented beyond using 300.
eta (simulated learning rate) = 1e-3 or 1e-4 per task (Table 5)
Controls the magnitude of the target SGD step in Lremove and is tuned per experiment.
gamma (repulsion strength) = not reported
Controls how strongly epsilon_target repels from the forbidden concept; value never stated.
lambda_remove, lambda_retain = not reported
Weights of the two loss terms in Lfinal; not reported in the main text.
LoRA rank = 1, 1, 4, 6 per task (Table 5)
Rank of generated LoRA matrices, varied per experiment without ablation.
Hypernetwork architecture = MLP, width/depth unspecified
Capacity is an implicit hand-chosen free parameter; no architecture details are given.

assumptions (4)

domain assumption Hypernet Fields gradient-matching is sufficient to train a hypernetwork without ground-truth weights.
Borrowed from Hedlin et al. 2025 and used as the backbone of Lremove; its reliability for unlearning is not independently validated.
domain assumption UnGuide's guided-regression loss is a valid unlearning task.
Ltask is adapted from a same-group prior work without independent validation on these benchmarks.
domain assumption Zero LoRA output for non-target concepts preserves generation quality.
Lretain penalizes deviation from zero; assumes this is sufficient to avoid catastrophic forgetting.
domain assumption CLIP text embeddings (or NV-Embed-v2 for celebrities) are a sufficient conditioning signal for the hypernetwork.
Central to the zero-shot claim; no explicit test of embedding-space coverage is provided.

invented entities (1)

CLIP-conditioned hypernetwork Hphi(c,s)
purpose: Generates LoRA weight changes for any concept c at trajectory step s
It is a trainable component; the paper provides no evidence that it generalizes to held-out concepts, and the synonym test overlaps with training.

how reviews work

0 comments

Cite this review

Pith. "Pith review of UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning." pith.science (2026). https://pith.science/paper/DY3AVV7V

@misc{pith2026260203410,
  author       = {Pith},
  title        = {Pith review of: UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DY3AVV7V}},
  note         = {Machine review of arXiv:2602.03410}
}

read the original abstract

Recent advances in large-scale diffusion models have intensified concerns about their potential misuse, particularly in generating realistic yet harmful or socially disruptive content. This challenge has spurred growing interest in effective machine unlearning, the process of selectively removing specific knowledge or concepts from a model without compromising its overall generative capabilities. Among various approaches, Low-Rank Adaptation (LoRA) has emerged as an effective and efficient method for fine-tuning models toward targeted unlearning. However, LoRA-based methods often exhibit limited adaptability to concept semantics and struggle to balance removing closely related concepts with maintaining generalization across broader meanings. Moreover, these methods face scalability challenges when multiple concepts must be erased simultaneously. To address these limitations, we introduce UnHype, a framework that incorporates hypernetworks into single- and multi-concept LoRA training. The proposed architecture can be directly plugged into Stable Diffusion as well as modern flow-based text-to-image models, where it demonstrates stable training behavior and effective concept control. During inference, the hypernetwork dynamically generates adaptive LoRA weights based on the CLIP embedding, enabling more context-aware, scalable unlearning. We evaluate UnHype across several challenging tasks, including object erasure, celebrity erasure, and explicit content removal, demonstrating its effectiveness and versatility. See the code on GitHub: https://github.com/gmum/UnHype.

Figures

Figures reproduced from arXiv: 2602.03410 by the authors.

**Figure 1.** Left: Comparative evaluation of explicit content erasure on the Flux architecture. We display the output of the original model alongside results from existing baseline methods and UnHype. Right: A parallel comparison conducted on Stable Diffusion, contrasting the original model’s generation against competing approaches and our proposed framework. Abstract Recent advances in large-scale diffusion models have intensi… view at source ↗

**Figure 2.** Overview of the inference in UnHype. The top part shows how the model handles an unlearned concept ("a photo of a cat"). The text embedding c is fed into a Hypernetwork that generates concept-specific LoRA parameters θS. These parameters modify the denoising model to suppress the forbidden concept, producing an alternative image (a forest) instead. The bottom part shows a retained concept ("a photo of a dog"). In th… view at source ↗

**Figure 3.** Overview of the removal loss in UnHype. The hypernetwork is queried at two consecutive steps, s and s + 1, to predict LoRA weights θs and θs+1. The difference between these weights, θs+1 − θs, forms the predicted step. Simultaneously, the target step of the task loss, ∆θtask, is computed according to Equation (5). The removal loss is the MSE Loss between the predicted step and the target step, forcing the hypernetwo… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Qualitative comparison showing object erasure results on Stable Diffusion, where the concept bird is mapped to a neutral concept. a fundamental obstacle: to learn a mapping of the form Hϕ(c) → θc (where θc denotes the LoRA weights corresponding to the concept c), one …

**Figure 5.** Figure 5: Qualitative comparison showing nudity erasure results on Flux. Prompts sampled from the I2P dataset. SD 1.4 ESD UCE MACE UnHype [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Qualitative comparison showing nudity erasure results on Stable Diffusion. Prompts sampled from the I2P dataset. Retention Loss (Lretain) This loss enforces the "semantic switch" behavior and prevents catastrophic forgetting. When Hϕ is conditioned on a "retain" concep…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 11 linked inside Pith

[2]

Salun: Empowering machine unlearning via gradient- based weight saliency in both image classification and generation.arXiv preprint arXiv:2310.12508,

Fan, C., Liu, J., Zhang, Y ., Wong, E., Wei, D., and Liu, S. Salun: Empowering machine unlearning via gradient- based weight saliency in both image classification and generation.arXiv preprint arXiv:2310.12508,

arXiv
[4]

C.2. Nudity Erasure Overview.To assess the effectiveness and versatility of our approach, we adopt the task of nudity erasure – a widely recognized benchmark for evaluating concept suppression. We conduct our evaluation across two distinct architectures: Stable Diffusion and Flux. Suppression Performance.Following established protocols, we first measure s...

2023
[5]

and Salimans, T

Ho, J. and Salimans, T. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598,

arXiv
[7]

Towards unbounded machine unlearning.Advances in neural information processing systems, 36:1957–1987,

Kurmanji, M., Triantafillou, P., Hayes, J., and Triantafillou, E. Towards unbounded machine unlearning.Advances in neural information processing systems, 36:1957–1987,

1957
[8]

K., Ruder, S., Dehghani, M., and Hender- son, J

Mahabadi, R. K., Ruder, S., Dehghani, M., and Hender- son, J. Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks.arXiv preprint arXiv:2106.04489,

arXiv
[10]

Unguide: Learning to forget with lora-guided diffusion models.arXiv preprint arXiv:2508.05755,

Polowczyk, A., Polowczyk, A., Malarz, D., Kasymov, A., Mazur, M., Tabor, J., and Spurek, P. Unguide: Learning to forget with lora-guided diffusion models.arXiv preprint arXiv:2508.05755,

arXiv
[11]

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

9 UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning Rando, J., Paleka, D., Lindner, D., Heim, L., and Tramèr, F. Red-teaming the stable diffusion safety filter.arXiv preprint arXiv:2210.04610,

arXiv
[12]

Semu: Singular value decompo- sition for efficient machine unlearning.arXiv preprint arXiv:2502.07587,

Sendera, M., Struski, L., Ksiazek, K., Musiol, K., Tabor, J., and Rymarczyk, D. Semu: Singular value decompo- sition for efficient machine unlearning.arXiv preprint arXiv:2502.07587,

arXiv

Show all 16 references

[13]

Erasediff: Eras- ing data influence in diffusion models.arXiv preprint arXiv:2401.05779,

Wu, J., Le, T., Hayat, M., and Harandi, M. Erasediff: Eras- ing data influence in diffusion models.arXiv preprint arXiv:2401.05779,

arXiv
[14]

a photo of the {erased class}

10 UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning A. Prompts used for training the hypernetwork We present the prompt sets used to train the hypernetwork across various tasks. For category erasure, we selected semantically related terms that do not appear in the...

2024
[2019]

a photo of {name}

with a confidence threshold of 0.6. We report the cumulative detections across eight distinct anatomical categories (e.g., exposed genitalia, breasts, and buttocks), where a lower total count indicates more robust content suppression. Our method clearly outperforms Flux-adapte...

2025
[2021]

Stable diffusion 1 vs 2-what you need to know

O’Connor, R. Stable diffusion 1 vs 2-what you need to know. Developer Educator at AssemblyAI.(Dec. 2022),[Online]. Available: https://www. assemblyai. com/blog/stable- diffusion-1-vs-2-what-you-need-to-know,

2022
[2022]

Kingma, D. P. and Welling, M. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

arXiv
[2023]

H., Chechik, G., and Cohen-Or, D

Gal, R., Alaluf, Y ., Atzmon, Y ., Patashnik, O., Bermano, A. H., Chechik, G., and Cohen-Or, D. An image is worth one word: Personalizing text-to-image generation using textual inversion.arXiv preprint arXiv:2208.01618,

arXiv
[2024]

Ha, D., Dai, A., and Le, Q

ECCV 2024 accepted. Ha, D., Dai, A., and Le, Q. V . Hypernetworks.arXiv preprint arXiv:1609.09106,

2024 arXiv
[2025]

and Deja, K

Cywi´nski, B. and Deja, K. Saeuron: Interpretable concept unlearning in diffusion models with sparse autoencoders. arXiv preprint arXiv:2501.18052,

arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [2]

Salun: Empowering machine unlearning via gradient- based weight saliency in both image classification and generation.arXiv preprint arXiv:2310.12508,

Fan, C., Liu, J., Zhang, Y ., Wong, E., Wei, D., and Liu, S. Salun: Empowering machine unlearning via gradient- based weight saliency in both image classification and generation.arXiv preprint arXiv:2310.12508,

arXiv

[2] [4]

C.2. Nudity Erasure Overview.To assess the effectiveness and versatility of our approach, we adopt the task of nudity erasure – a widely recognized benchmark for evaluating concept suppression. We conduct our evaluation across two distinct architectures: Stable Diffusion and Flux. Suppression Performance.Following established protocols, we first measure s...

2023

[3] [5]

and Salimans, T

Ho, J. and Salimans, T. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598,

arXiv

[4] [7]

Towards unbounded machine unlearning.Advances in neural information processing systems, 36:1957–1987,

Kurmanji, M., Triantafillou, P., Hayes, J., and Triantafillou, E. Towards unbounded machine unlearning.Advances in neural information processing systems, 36:1957–1987,

1957

[5] [8]

K., Ruder, S., Dehghani, M., and Hender- son, J

Mahabadi, R. K., Ruder, S., Dehghani, M., and Hender- son, J. Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks.arXiv preprint arXiv:2106.04489,

arXiv

[6] [10]

Unguide: Learning to forget with lora-guided diffusion models.arXiv preprint arXiv:2508.05755,

Polowczyk, A., Polowczyk, A., Malarz, D., Kasymov, A., Mazur, M., Tabor, J., and Spurek, P. Unguide: Learning to forget with lora-guided diffusion models.arXiv preprint arXiv:2508.05755,

arXiv

[7] [11]

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

9 UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning Rando, J., Paleka, D., Lindner, D., Heim, L., and Tramèr, F. Red-teaming the stable diffusion safety filter.arXiv preprint arXiv:2210.04610,

arXiv

[8] [12]

Semu: Singular value decompo- sition for efficient machine unlearning.arXiv preprint arXiv:2502.07587,

Sendera, M., Struski, L., Ksiazek, K., Musiol, K., Tabor, J., and Rymarczyk, D. Semu: Singular value decompo- sition for efficient machine unlearning.arXiv preprint arXiv:2502.07587,

arXiv