acceptodds
Withdrawn from ICLR 2027

Filtering Memorization from Parameter-Space in Diffusion Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Low-Rank Adaptation (LoRA) enables efficient customization of diffusion models, but may also memorize and reproduce sensitive or copyrighted training content. This risk is particularly relevant in LoRA-sharing ecosystems, where adapters are distributed without their training data or training pipelines. Existing mitigation methods typically require access to training data, intervention during training, or control over inference, making them unsuitable when only a public pretrained backbone and a third-party LoRA are available. We propose Base-Anchored Filtering (BAF), a training- and data-free framework for post-hoc memorization mitigation. BAF spectrally decomposes the LoRA update and measures each component's alignment with the pretrained backbone's principal subspace, preserving strongly anchored components while suppressing weakly anchored components associated with instance-specific memorization. Experiments across three datasets, three diffusion backbones, and 20 community-released LoRAs show that BAF consistently reduces memorization while preserving generation quality, demonstrating that pretrained parameter geometry provides a useful structural prior for filtering memorization from released LoRAs. Our code is in the supplementary material.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Diffusion Models, Memorization, Low-Rank Adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_kH56OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors analyze memorization in LoRA finetuning of pretrained diffusion models and provide an SVD-based mitigation method.

Specifically, they apply SVD to the LoRA updates, and find that the singular vectors that do not align with the top singular subspaces of the pretrained weights are related to memorization. So they propose to prune or reweight such components and find that doing so reduces the SSCD similarity between generated images and the training images. To validate their method, the authors conduct experiments on different architectures (SD1.5, SDXL, and FLUX1.) and different datasets (PokenMon, CelebA, and LAION-ART).

Critical strengths
  1. The authors study a practical and meaningful setting, that is the memorization in the LoRA finetuning of pretrained diffusion models, where the downstream data can be scarce and memorization can be severe.
  2. The method is stated clearly and compared against a reasonably wide range of other methods.
Critical weaknesses
  1. The method and observations lack justification, and I do not intuitively get why alignment with the top singular subspaces with the pretrained weights matters for memorization.
  2. The theory is largely decorative and not very relevant, and the core ideas that need to be justified (alignment ↔\leftrightarrow generalization) are directly used as an assumption.
  3. The evaluation is not entirely reasonable, where the major improvement is reducing the SSCD (while a reasonably high SSCD is partly expected for getting closer to the finetuning distribution). The other two metrics (CLIP and aesthetic score) do not directly reflect the effectiveness of finetuning, which is aimed at generating a specific distribution of downstream images rather than improving generic CLIP/aesthetic scores. Also, CLIP score is quite old, and the authors may test their method on more modern metrics like GenEval.
Questions for authors
  1. For Fig.2, it would be better to include some intuitive examples as in Fig.1.
  2. From Fig.3, it seems that real-world LoRA without mitigation is not memorizing? And also BAF seems not substantially better than other methods like RTA.
Reviewer_LaokOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses memorization in diffusion LoRAs in a weight-only setting, that is, where a third-party adapter is released without training data or control over inference. It proposes Base-Anchored Filtering (BAF). BAF decomposes each LoRA update into rank-1 spectral channels via SVD. It scores each channel by its bilateral alignment with the top-K principal subspaces of the pretrained weight, choosing K per layer with an energy-ratio rule. It then removes (BAF-Hard) or attenuates (BAF-Soft) channels that score below a random-projection baseline,. The paper presents a recovery bound relative to a counterfactual "clean" LoRA. Validation covers controlled memorization (50 canaries duplicated 50×) on Pokemon, CelebA-HQ, and LAION-Art with SD1.5, SDXL, and Flux.1, plus a 10-rater blind study on 20 CivitAI LoRAs. BAF reduces targeted SSCD relative to the inference-time baselines while maintaining or improving CLIP and aesthetic scores.

Critical strengths

Practical and well-motivated problem setting. The weight-only, data-free setting matches how LoRAs are actually distributed, and it is a setting where existing training-time, inference-time, and concept-erasure defenses do not apply. The method is simple, needs no training, adds no inference cost, and the backbone SVD can be cached across adapters. This is the main reason the work is relevant in practical settings.

Consistent empirical trend. Across three datasets and three backbones, BAF lowers SSCD relative to the unfiltered LoRA and the inference-time baselines without degrading quality metrics. The directional ablation (high-align-only vs. low-align-only) gives intuitive support for the anchoring hypothesis.

Critical weaknesses

Missing controls: it is unclear that the anchoring score itself is what drives the gains. The central claim is that alignment with the pretrained principal subspace separates memorization from generalizable adaptation. The paper does not compare against simpler parameter-space controls that need equally little information: rank truncation (keeping the top-k singular channels of delta-W), singular-value thresholding, global LoRA scaling, or removing random channels. This matters because prior work (Gu etal) links memorization strength to LoRA rank. Without these controls, the reported gains could come from generic rank or energy reduction rather than base anchoring, and the paper's main conceptual contribution will not be established. The fact that CLIP and aesthetic scores often improve after filtering also points to a regularization effect that needs to be separated from the proposed mechanism.

Secondly, The core hypothesis and the theory have limited support, and real-world evaluation is weak. The theory is validated only in one setting (SD1.5/Pokemon), and the measured K = 0.45 means nearly half of the memorization energy lies outside what channel-wise filtering can reach. That is hard to reconcile with the claim that the separation conditions are "satisfied." The behavioral corollary rests on a local-Lipschitz assumption that is not tested. The controlled benchmark uses extreme duplication (50 times), which is unrepresentative of natural memorization. The real-world evidence is subjective only: 10 raters, with the high/low-memorization grouping assigned by eye and no objective measurement. The hyperparameters also differ between the stated defaults (alpha = 1, tau = 0.85) and the qualitative figures (alpha = 0.4, tau = 0.8), which raises concerns about tuning for each figure. Together, these weaken the claim that BAF works reliably on LoRAs found in practice.

Questions for authors

How does BAF compare with rank truncation, singular-value thresholding, global LoRA down-scaling, and random channel removal, each matched to the same fraction of removed energy, on SSCD and quality?

How many channels, and what fraction of ΔW energy, are removed per layer on average? Does BAF's advantage hold at matched removed energy?

Can the values be be reported for at least one more backbone and dataset? How do the authors reconcile k = 0.45 with the stated claim that the separation conditions hold?

Were alpha and tau's fixed across all quantitative and qualitative results? If not, how were they selected, and without access to training data?

For the CivitAI study, how were memorized identities verified, and can an objective measure be reported (e.g., similarity to reference images of the known subject)?

Reviewer_jW9rOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper targets a weight-only setting for diffusion LoRAs: a hosting platform or downstream user has the released adapter and the public base model, but no training data, no training pipeline, and no control over inference. It proposes Base-Anchored Filtering (BAF): SVD the LoRA update ΔW into rank-1 spectral channels, score each channel by its bilateral alignment with the top-K principal subspace of the corresponding pretrained weight (a_i = ‖P_L u_i‖²·‖P_R v_i‖², Eq. 1), and either drop channels below a random-projection baseline a_null = K²/(mn) (BAF-Hard) or scale each channel by a_i^α (BAF-Soft). K is chosen per layer by a cumulative-energy rule (Eq. 4, τ_energy = 0.85). The claim is that strongly anchored channels carry generalizable adaptation while weakly anchored channels carry instance-specific memorization. Evidence: a directional ablation on SD1.5 (Fig. 2); controlled canary-duplication experiments (50 canaries × 50 duplicates) on Pokémon, CelebA-HQ and LAION-Art across SD1.5/SD2.1/SDXL/Flux.1 with targeted SSCD, CLIP score and aesthetic score against training-time (dedup, DP-LoRA) and inference-time (RTA, prompt optimization, CFG-Reverse, AMG) baselines (Tab. 1–2, App. D); a qualitative study plus 10-rater blind evaluation on 20 Civitai LoRAs (Fig. 3, App. E); a recovery bound in terms of leakage coefficients measured on one SD1.5/Pokémon pair (Thm. 1–2, Tab. 3); and ablations over α, τ_energy and duplication factor (App. F).

Critical strengths
  1. A realistic and under-served threat model with a genuinely data-free, training-free, inference-agnostic mitigation (Sec. 3 "Problem Setup", Sec. 4.2, App. G). The weight-only setting matches how adapters are actually distributed on Civitai/Hugging Face, and BAF needs nothing beyond the adapter and the public backbone; it is a one-time offline edit (38 s on SD1.5, 343 s on SDXL, cacheable per backbone) with zero inference overhead. This is consequential because every compared baseline either needs the training set or control of sampling, so if the method's premise holds it is the only option available to a hosting platform.
  2. Breadth of the empirical program (Tab. 1–2, Tab. 4–9, App. E–F). Three datasets, four backbones including Flux.1, seven baselines, three memorization and four utility metrics, 20 real community LoRAs with a blind user study, and sensitivity analyses over α, τ_energy and duplication factor. The consistent direction of the effect across all of these makes it unlikely that the reported SSCD reductions are an artifact of one configuration.
Critical weaknesses
  1. The central geometric claim has no control against plain shrinkage of ΔW, and the experimental regime makes shrinkage alone a sufficient explanation (Sec. 4.1, Fig. 2, Tab. 1–2, App. A/D). BAF-Soft multiplies every channel by a_i^α ∈ [0,1] and BAF-Hard discards a subset of channels; both reduce ‖ΔW‖. In the controlled setup, Pokémon has 833 training images while the 50 canaries are duplicated 50× (≈2,500 images), so roughly three quarters of the training set is canary content and the "High-mem" LoRA is largely an overfit to 50 images. In that regime any attenuation of the update — a uniform scale s·ΔW, truncation to the top-k channels by σ_i, or random channel dropping at matched retained energy — is expected to lower targeted SSCD and may raise AES/CLIP, which is exactly the pattern reported for BAF (utility improves over the unfiltered LoRA in most cells of Tab. 1–2). The directional ablation in Fig. 2 does not resolve this: the low-align-only variant retains fewer and weaker channels, so its quality collapse is expected regardless of what those channels encode. Without a shrinkage/truncation/random-subset control at matched energy, Contribution 1 ("spectral alignment separates generalizable adaptation from memorization") is not established, and the comparison to inference-time baselines does not substitute for it. This is fixable in a rebuttal, but it determines whether the paper's mechanism is real.
  2. The released code implements a different method from the one described, and the real-world evaluation conflates a LoRA's intended subject with memorization (Sec. 4.3, Eq. 4, Alg. 1; baf_mitigation.py; Sec. 5.2, Fig. 3, App. E.1). In the supplementary code, K is int(0.5·min(m,n)) (fixed base_rank_ratio), not the energy-adaptive rule of Eq. 4 that Sec. 4.3 presents as the architecture-aware contribution and that App. F analyzes (Fig. 9–10); the hard gate uses a fixed threshold=0.3, not a_null; the soft default is power=2.0, not α=1; and the README defines nL2 as "lower is better / more negative = less memorized" while the paper's Tab. 4–9 report positive nL2 with "higher is better", indicating the tables were produced with a different evaluation script. Either the reported numbers were not produced by the released method or Sec. 4.3 was not used; both are soundness and reproducibility problems. Separately, for subject/character adapters the "memorized identity" is the concept the adapter exists to produce: in Fig. 3 the Goku toy LoRA after BAF generates a generic blond figure in toy style, i.e., the method acts as concept erasure, and the high-/low-memorization categorization in App. E.1 is defined by "consistently reproduce a recognizable identity." From weights alone BAF cannot distinguish a verbatim training instance from the learned subject, so the method's scope should be restricted to style/domain LoRAs or evidence should be given that it preserves the subject while removing the instance. Supporting issues: no clean (no-duplication) LoRA row in Tab. 1–2 as the SSCD floor and the reference for Thm. 2; the number of generated samples and the prompt set for CLIP/AES are not stated anywhere; no seeds or confidence intervals; post-BAF targeted SSCD remains ≥0.5 in most SDXL/Flux cells (e.g., 0.68 CelebA-SDXL, 0.62 Pokémon-SDXL) against the "memorization-free" wording of Fig. 1; and the measured leakage quantities (ρ_mem = 0.29, ρ_clean = 0.22, κ = 0.45, Tab. 3) give a recovery bound of ≈0.74‖ΔW_mem‖ + 0.22‖ΔW_clean‖, which does not support "approximately recovers the counterfactual clean LoRA."
Questions for authors

Clarification requests (answers could change my assessment):

  1. Please add the following controls to Tab. 1 (at least SD1.5 and SDXL on Pokémon and CelebA-HQ), each at retained Frobenius energy matched to BAF-Soft and BAF-Hard: (a) uniform scaling s·ΔW; (b) truncation to the top-k channels by σ_i; (c) random channel subset. Does BAF still improve SSCD at equal utility, and by how much?
  2. Which K-selection and thresholds produced Tab. 1–2 and Fig. 9–10: the energy-adaptive rule of Eq. 4 with τ_energy = 0.85 and a_null, or the fixed base_rank_ratio = 0.5 / threshold = 0.3 / power = 2.0 in the released baf_mitigation.py? Please reconcile the code and the paper, including the sign convention of nL2.
  3. Please add a clean-LoRA (no duplication) row to Tab. 1–2 for every backbone, and state the number of generated images, the prompt set, seeds, and standard deviations used for SSCD/CLIP/AES.
  4. For subject or character LoRAs (e.g., the Goku adapter in Fig. 3), what is the intended behavior: should the subject survive filtering? If yes, please show a case where BAF keeps the subject but removes a duplicated training instance; if no, please state that BAF is a concept-removal method for such adapters and restrict the claims to style/domain adapters.
  5. What fraction of the training set is canary content in each controlled setting (Pokémon 833 images vs. 50×50 duplicates), and do the conclusions hold at a realistic duplication level (e.g., D ≤ 5 with Nc = 50 on LAION-Art)? Fig. 11 reports targeted SSCD only; please add utility at each D.

Suggestions (not affecting the rating): report AES/CLIP of the pretrained base model as a reference so that "utility improves after BAF" can be read against regression to the base; fix "Assumption 1" being used for two different assumptions (Sec. 4.4 vs. App. C.4) and "Table 3 to Table 8" (App. D.1, should be Tab. 4–9).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.