acceptodds
Under review as a conference paper at ICLR 2027

LoRA Scope: Semantic Change Probing for LoRA Search and Evaluation

Abstract

The rapid growth of LoRA adapters for text-to-image diffusion models has made them difficult to understand, evaluate, and search beyond thumbnails and noisy metadata. We introduce LoRA Scope, a framework that represents each LoRA by what it changes: the CLIP-space difference between matched base and LoRA-adapted generations. The framework connects three complementary components through diagnostic prompts and generated images. LoUPE uses VLM-guided prompt exploration to construct adapter-specific diagnostic prompts; the LoUPE Score measures the stability, precision, and strength of the induced changes and aligns with human judgments; and LoCAL aligns LoRA weight representations with CLIP embeddings of these prompts and LoRA-generated images through learned query adapters, enabling text- and image-to-LoRA retrieval without query-time generation. On a large corpus of real-world LoRAs, LoRA Scope provides interpretable adapter characterization, human-aligned quality assessment, and scalable metadata-free retrieval directly from LoRA weights.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Low-Rank Adaptation (LoRA), Text-to-Image Diffusion Models, Multimodal Model Retrieval, Weight-Space Representation Learning, Adapter Evaluation, Semantic Change Probing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_KbyZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The search and management of existing LoRA communities mainly rely on titles, tags, descriptions, and example images, which may not accurately reflect the actual abilities learned by LoRA. To this end, the paper proposes the LoRA Scope framework, which advocates analyzing the semantic changes generated between the base model and the LoRA adaptation model, characterizing the semantic changes actually introduced by LoRA, and further learning to directly predict its semantic representation from LoRA weights. The LoRA Scope consists of three modules: LoUPE, LoUPE Score, and LoCAL, which correspond to the understanding, evaluation, and search of LoRA research.

Critical strengths
  1. With the growing popularity of personalized image generation, community LoRA libraries contain a large number of adapters with different functions and quality levels. How to understand its actual ability, evaluate its reliability, and conduct efficient retrieval is a problem with clear practical value. The research problem has clear application requirements.

  2. LoUPE uses LLMs to expand prompts and monitors changes in CLIP-space statistics to determine when to stop. It also uses VLM feedback to replace prompts that fail to activate the adapter effectively. Has a certain degree of innovation.

  3. The author provided sufficient explanations for each module, and the paper is clear and easy to understand

  4. The author conducted large-scale experiments and conducted valuable manual evaluations

Critical weaknesses
  1. LoUPE stops Prompt extension based on two conditions:The cumulative variance curve explained by adjacent rounds of PCA has a sufficiently small change;The cosine similarity in the direction of average difference between adjacent wheels is sufficiently high. These conditions can measure the stability of a specific statistic, but are not sufficient to ensure the convergence of a complete difference distribution.From an experimental perspective, Figure 4 mainly demonstrates that the stopping conditions designed by the author can be satisfied within a finite number of rounds, but does not fully demonstrate that these Prompts have accurately covered the actual capabilities of LoRA.
  2. This article lacks a comparison between LoUPE's adaptive prompt exploration and fixed prompt, random prompt extension, and LLM only fixed round extension under the same generation budget.
  3. The three factor product in Figure 6 did not show a higher correlation with the overall score than when Precision or Strength were used alone. Therefore, the paper has not fully demonstrated the necessity of using a multiplication combination of three indicators.
  4. This article mainly compares LoCAL with retrieval methods such as Stylus, Stylus+, and Stylus++, but lacks comparison with other weight retrieval baselines.
Questions for authors

see weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer KbyZ (1/3)

We thank the reviewer for the constructive suggestions on validating prompt exploration and strengthening the evaluation. Below, we address the concerns and questions raised in the review.

W1: Stopping conditions, convergence, and coverage of actual capabilities

Weakness 1: LoUPE stops Prompt extension based on two conditions:The cumulative variance curve explained by adjacent rounds of PCA has a sufficiently small change;The cosine similarity in the direction of average difference between adjacent wheels is sufficiently high. These conditions can measure the stability of a specific statistic, but are not sufficient to ensure the convergence of a complete difference distribution.From an experimental perspective, Figure 4 mainly demonstrates that the stopping conditions designed by the author can be satisfied within a finite number of rounds, but does not fully demonstrate that these Prompts have accurately covered the actual capabilities of LoRA.

We tested whether LoUPE's stopping rule yields reproducible exploration and broad coverage of observed adapter behavior through two additional experiments. With LoUPE initialized from informative community prompts, independent runs show substantial overlap in adapter-effect space, and the final prompt sets cover 87% of held-out community usage.

  • Agreement across runs. We independently ran LoUPE twice on each of 30 adapters and compared the final prompt sets through nearest-neighbor matching in adapter-effect space (δ_II). On average, 84.5% of prompts in one run have an effect within the nearest-neighbor radius of an effect in the other run. The corresponding overlap is 89.0% between random halves of the same run and 3% between different adapters. The mean effect directions also have cosine similarity 0.925. These results support reproducibility of the explored effects across independent runs.

  • Coverage of held-out usage. For the 92 adapters with at least four distinct community prompts, we held out a subset that LoUPE never saw. Coverage measures the fraction of held-out prompts whose adapter effect δ_II lies within a nearest-neighbor radius of an effect in the explored set. At equal prompt counts, LoUPE covers 87% of held-out usage, compared with 43% for a fixed generic prompt list from PartiPrompts. Coverage increases from 64% after the initial round to 81% halfway through exploration and 87% at the stopping round. The final round adds only 1.2 percentage points, providing empirical evidence of diminishing coverage gains near the stopping point.

In the revision, we will describe the stopping rule as an empirical saturation criterion and report these tests as evidence of reproducibility across runs and substantial coverage of observed usage. Held-out community prompts provide a proxy for realized adapter behavior. We will make clear that these tests do not establish convergence of the full difference distribution or exhaustive coverage of all capabilities, and that the evaluation uses informative starting examples.

AuthorsOpenReview ↗

Response to Reviewer KbyZ (2/3)

W2: Prompt-exploration baselines under the same generation budget

Weakness 2: This article lacks a comparison between LoUPE's adaptive prompt exploration and fixed prompt, random prompt extension, and LLM only fixed round extension under the same generation budget.

We compared LoUPE with fixed prompts, random prompt extension, and LLM-only extension on 100 randomly sampled adapters. With informative community prompts as starting examples, LoUPE covers more held-out usage than fixed or randomly extended prompt sets at matched prompt counts. The LLM-only comparisons further distinguish the coverage gained through prompt exploration from the activation gained through image-grounded refinement.

Coverage was evaluated on the 92 adapters with at least four distinct community prompts. We held out a subset of these prompts that was never shown to any method. Coverage is the fraction of held-out prompts whose adapter effect δ_II lies within a nearest-neighbor radius of an effect in the evaluated prompt set. Activation measures the magnitude of the adapter's effect as 1 − cos(base image, LoRA image) in CLIP space, averaged over the prompt set. LoUPE, the LLM-only variants, and random extension use the same community prompts as starting examples. The table reports the full prompt sets, with prompt counts averaged across adapters. The matched-count comparison with fixed prompts is reported below.

Prompt setMean prompts per adapterCoverage of held-out usageActivation
Fixed prompts (PartiPrompts)1000.470.12
Random extension with PartiPrompts650.700.13
LLM-only, initial round80.64—
LLM-only, half of LoUPE's rounds340.81—
LLM-only, same rounds as LoUPE650.870.21
LoUPE650.870.22
  • Fixed prompts. We use the same 100 prompts, sampled at random from PartiPrompts, for every adapter. This fixed list covers 47% of held-out usage, compared with 87% for LoUPE using 65 prompts on average. After subsampling both sets to equal prompt counts for each adapter, averaging 58 prompts, coverage is 0.87 for LoUPE and 0.43 for the fixed list (difference 0.44, 95% CI [0.37, 0.50]). Mean activation is also higher for LoUPE, at 0.22 compared with 0.12 for fixed prompts.

  • Random extension. We extend LoUPE's initial round of 8 generated prompts with prompts sampled at random from the PartiPrompts list, matching the size of LoUPE's final set for each adapter. Results are averaged over 20 random draws. Random extension reaches coverage of 0.70, compared with 0.87 for LoUPE (difference 0.17, 95% CI [0.13, 0.22]), with mean activation of 0.13 compared with 0.22. Starting from the same initial coverage of 0.64, adaptive exploration therefore gains more coverage than random extension at the same prompt count.

  • LLM-only extension. We evaluate LLM-only prompt sets after the initial round, halfway through LoUPE's rounds, and at LoUPE's stopping round for each adapter. Coverage increases from 0.64 to 0.81 and 0.87, respectively. The final round adds only 1.2 percentage points. LoUPE stops between iterations 5 and 12, averaging 7.1 across adapters. These results support the stopping rule as an empirical saturation criterion that adapts the number of rounds to each adapter.

  • VLM refinement. After the same number of LLM rounds, the VLM stage increases mean activation from 0.21 to 0.22 while preserving coverage at 0.87. It replaces about 8 of the 65 prompts per adapter. The replacements have 1.7 times the activation of the prompts they replace (0.225 compared with 0.129, difference 0.096, 95% CI [0.088, 0.104]), showing that image-grounded refinement identifies prompts that elicit stronger adapter effects.

We will add these baseline comparisons and their prompt-count protocols to the revised experiments. Consistent with the evaluation scope discussed above, these results concern exploration from informative starting examples and coverage of held-out community usage.

AuthorsOpenReview ↗

Response to Reviewer KbyZ (3/3)

W3: Necessity of the multiplicative LoUPE Score

Weakness 3: The three factor product in Figure 6 did not show a higher correlation with the overall score than when Precision or Strength were used alone. Therefore, the paper has not fully demonstrated the necessity of using a multiplication combination of three indicators.

We use multiplication because a high LoUPE Score should require stable, precise, and sufficiently strong interventions jointly. A low value in any component suppresses the product, even when the others score highly. The correlation comparison in Fig. 6 does not by itself establish that Precision or Strength alone can replace this joint quality criterion.

Appendix C.2 (Fig. 10) illustrates their distinct failure modes. Low Stability corresponds to inconsistent effects across generations, low Precision to imprecise modifications, and low Strength to barely perceptible interventions. These examples motivate evaluating all three dimensions jointly.

W4: Other weight-based retrieval baselines

Weakness 4: This article mainly compares LoCAL with retrieval methods such as Stylus, Stylus+, and Stylus++, but lacks comparison with other weight retrieval baselines.

Our setting retrieves community LoRAs from text or image queries using candidate weights alone. Weight-learning methods such as ProbeX / Tree Experts and Learning on LoRAs (LoL) are relevant, but their published evaluations do not provide directly comparable results under this retrieval protocol. LoL's GL-equivariant encoder can be adapted to our framework with minimal changes, allowing us to compare weight encoders within the proposed setting.

We therefore conducted an encoder ablation by replacing LoCAL's W2T-based encoder backbone with LoL's GL-equivariant architecture, retaining the same SD1.5 collection, behavioral supervision, and contrastive retrieval objective. Both encoders are evaluated on the same 1,254 held-out test LoRAs, with Recall@K and MRR averaged across LoRAs.

Weight encoderQuery modalityCandidatesR@1 (%)R@5 (%)R@10 (%)MRR
GL-equivariantText1,2541.254.427.330.0368
LoCALText1,2544.0711.9017.490.0885
GL-equivariantImage1,2543.059.3914.110.0701
LoCALImage1,2546.0316.2122.920.1184

This ablation shows that encoder choice matters for retrieval and supports the effectiveness of LoCAL's representation in the studied setting.

Dear Reviewer,

The discussion period is nearing its end, and we haven’t yet seen a reply from you. We encourage you to respond to the authors’ comments, especially if you have further questions about points that may affect your assessment.

Thank you for taking the time to engage in the discussion.

Best,

Your AC

Reviewer_hEMxOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LoRA Scope, a framework for understanding, evaluating, and retrieving LoRA adapters for text-to-image diffusion models. Its core idea is to characterize each LoRA by the change it induces: the CLIP-space difference between images generated with and without the adapter from the same prompt. LoUPE uses an LLM and a VLM to iteratively build adapter-specific diagnostic prompts until the distribution of these difference vectors stabilizes, and the LoUPE Score rates adapter quality by the stability, precision, and strength of that distribution. LoCAL learns to map LoRA weights into CLIP space, enabling text- and image-to-LoRA retrieval from weights alone. Experiments on SD1.5 LoRAs show that the LoUPE Score aligns better with human judgments than image-level metrics, and that LoCAL is competitive with metadata-based retrieval baselines.

Critical strengths
  • The intervention-based view of evaluating an adapter is interesting. It compares LoRA outputs against matched base-model outputs instead of scoring the generated images in isolation. This separates “does the adapter learn a coherent change” from “are the images good,” a distinction existing image-level metrics miss, and the paper supports it with both correlation analyses and qualitative examples.
Critical weaknesses
  • Motivation. The central premise is that LoRA metadata is noisy and reflects creator intent rather than realized behavior, but it is not well supported.
    • The claim is never measured. How often is metadata wrong, missing, or misleading in the corpus? Without this, it is unclear how large the problem LoRA Scope solves actually is.
    • The method is not consistent with this motivation. LoUPE initializes its prompt pool from user prompts, trigger words, and metadata tags, and the Stage-1 LLM expands prompts from them. LoCAL is metadata-free only at inference time, because its training targets come from LoUPE outputs, which depend on metadata. If metadata is unreliable, the pipeline should not rely on it.
  • Limited generality of LoCAL. The LoCAL encoder assumes each layer update is a low-rank product of factor matrices (QR/SVD canonicalization, rank standardization, bilinear measurements over paired singular vectors). It therefore cannot handle other parameter-efficient fine-tuning methods that do not produce such weight deltas, such as Textual Inversion. The paper should state this scope explicitly.
Questions for authors
  • Could you provide results on more recent models, such as SDXL or DiT-based models (e.g., FLUX)?
  • Does LoUPE work with completely wrong or empty metadata? For example, how do the converged prompt sets, LoUPE Scores, and downstream LoCAL retrieval change when the prompt pool is initialized only from generic prompts?

(Minor)

  • The acronym LoUPE is never explicitly expanded; it is only implied by the title of Sec. 3.2. Please define it at first use.
6 comments
AuthorsOpenReview ↗

Response to Reviewer hEMx (1/3)

We thank the reviewer for the detailed feedback on our motivation, metadata assumptions, and the scope of our evaluation.

We clarify that our goal is LoRA model retrieval without candidate metadata at inference. To learn this capability, we use a curated, metadata-rich dataset where LoUPE explores beyond supplied examples to characterize adapter effects, providing supervision for LoCAL to learn retrieval semantics directly from weights.

We detail these clarifications and experiments below.

W1.1 Metadata quality

Weakness 1.1: Motivation. The central premise is that LoRA metadata is noisy and reflects creator intent rather than realized behavior, but it is not well supported.

  • The claim is never measured. How often is metadata wrong, missing, or misleading in the corpus? Without this, it is unclear how large the problem LoRA Scope solves actually is.

Finding LoRAs for a desired visual effect in large community libraries is difficult when their accompanying documentation is limited or unavailable. Audits of Hugging Face across model types found missing model cards for at least 36% of the models surveyed by Horwitz et al. and more than half of the repositories in the 2022 snapshot studied by Liang et al.. These documentation gaps motivate our goal of retrieving adapters directly from weights, without requiring candidate metadata or examples at inference.

To learn this retrieval capability, we use ImageGEM, a curated dataset with safety filtering and structured records linking LoRAs to descriptive metadata (titles, descriptions, tags, and trigger words) and example images with their paired prompts. This metadata-rich corpus supports the behavioral probing used to construct LoCAL's training supervision.

We also examine how informative each component of the curated metadata is, through three experiments on agreement of titles and descriptions with examples, tag and trigger-word distinctiveness, and prompt coverage. These studies draw on the 6,273 adapters preprocessed for LoCAL within our 6,635-adapter corpus (Sec. 4.1).

  • Descriptive text fields' agreement with examples. For each LoRA, we compare its title or description with its own prompts and gallery images using CLIP ViT-B/32. We repeat the comparison with the title or description unchanged, using examples from 50 randomly selected other LoRAs. The table reports mean cosine similarities, averaged within LoRAs and then across LoRAs.

    ComparisonSame LoRAWrong LoRA
    Title–original prompts0.5920.499
    Description–original prompts0.6520.583
    Description–original gallery images0.2630.220

    Titles and descriptions agree more closely with their own examples, and all three differences are statistically significant based on 95% confidence intervals.

  • Tag and trigger-word distinctiveness. We compare normalized tag sets and trigger-phrase sets across LoRA versions.

    MeasurementTagsTrigger phrases
    Versions sharing an identical set with another version31.79%6.65%
    Median number per version51

    Tags are frequently shared across versions, while trigger phrases are more distinctive but sparse.

  • Prompt coverage. We measure prompt coverage as the percentage of prompts with a counterpart in the other set at CLIP ViT-B/32 cosine similarity of at least 0.6. For each LoRA, we randomly select equal numbers of prompts from the uploader-provided and LoUPE-generated sets, up to 16 from each. We repeat this sampling 50 times and average coverage first for each LoRA, then across LoRAs. LoUPE-generated prompts cover 82.69% of uploader-provided prompts, while uploader-provided prompts cover 65.86% of LoUPE-generated prompts. This asymmetry supports broader exploration in prompt embedding space, without establishing exhaustive coverage of adapter capabilities.

In addition to these metadata studies, we test retrieval under swapped inputs among 274 selected candidate LoRAs. Exchanging gallery images and paired prompts between dissimilar LoRAs reduces Stylus++ text Recall@1 from 33.03% to 1.35%, even with correct descriptive fields. LoCAL remains at 11.07% because it indexes weights. This tests dependence on aligned examples, without estimating the prevalence of incorrect metadata. The full retrieval tables and protocol appear in our response to Reviewer j6YW (W2 & Q2).

Together, the three metadata studies motivate probing beyond descriptive fields and supplied examples, while the swap test shows the sensitivity of Stylus++ retrieval to mismatched examples. LoUPE uses community prompts to guide broader behavioral probing, and LoCAL learns from the resulting supervision to retrieve held-out adapters from weights alone, without candidate metadata or example images at inference.

AuthorsOpenReview ↗

Response to Reviewer hEMx (2/3)

W1.2 & Q2 LoUPE and metadata

Weakness 1.2:

  • The method is not consistent with this motivation. LoUPE initializes its prompt pool from user prompts, trigger words, and metadata tags, and the Stage-1 LLM expands prompts from them. LoCAL is metadata-free only at inference time, because its training targets come from LoUPE outputs, which depend on metadata. If metadata is unreliable, the pipeline should not rely on it.

Question 2: Does LoUPE work with completely wrong or empty metadata? For example, how do the converged prompt sets, LoUPE Scores, and downstream LoCAL retrieval change when the prompt pool is initialized only from generic prompts?

LoUPE and LoCAL serve complementary purposes. LoUPE uses available examples to guide behavioral probing, then expands and refines the prompt set through LLM exploration and image-grounded feedback to uncover capabilities that initial examples may not reveal. The resulting behavioral measurements can enrich an adapter's metadata and provide supervision for LoCAL to retrieve unseen adapters from weights alone. Informative documentation in a curated training corpus thus supports retrieval for adapters whose metadata or examples are unavailable at inference.

All 6,635 adapters in our corpus have community prompts, with a median of 21 distinct prompts per adapter. To test how initialization affects LoUPE's evaluation, we ran LoUPE on 100 randomly sampled adapters using gpt-6-luna as the Stage-1 LLM because the original o4-mini was no longer available for these experiments. We compare community prompts, cold-start initialization, and mismatched prompts from another adapter. In the cold-start setting, the LLM generates initial prompts without adapter-specific metadata or examples. We also evaluate a fixed PartiPrompts list without LoUPE. Coverage was evaluated on the 92 adapters with sufficient community prompts for a held-out split. The held-out prompts were never shown to LoUPE. Coverage is the fraction of held-out prompts whose adapter effect δ_II lies within a nearest-neighbor radius of some prompt in the final set. Score rank correlations are measured against the run initialized with community prompts.

Prompt setCoverage of held-out usageMean stopping iterationScore rank correlation
LoUPE with community prompts0.877.1—
LoUPE with cold-start initialization0.5310.80.04
LoUPE with mismatched prompts0.567.8−0.04
Fixed prompts from PartiPrompts0.47——
  • Final prompt sets and stopping behavior. With community prompts, LoUPE covers 87% of held-out usage. With cold-start initialization, LoUPE reaches 53% coverage, compared with 47% for the fixed PartiPrompts list (difference 0.06, 95% CI [0.03, 0.10]). In this setting, its VLM stage identifies replacement prompts with stronger adapter activation than those they replace (0.17 versus 0.03). Informative examples nevertheless make exploration more effective. Community prompts yield substantially higher coverage and earlier stopping than cold-start initialization (7.1 versus 10.8 iterations), while mismatched prompts yield 56% coverage at a mean stopping iteration of 7.8. Thus, LoUPE can explore without adapter-specific starting information, while relevant examples help direct that exploration toward observed adapter usage.
  • LoUPE Scores. An independent rerun initialized with community prompts yields a rank correlation of 0.73 with the reference run. Cold-start and mismatched initialization yield correlations of 0.04 and −0.04. Score rankings thus reflect the behavior explored by each prompt set. For evaluation intended to characterize community usage, informative adapter-specific examples provide a useful basis for exploration. The cold-start results do not establish that the resulting scores characterize the same usage.
  • LoCAL retrieval. LoCAL indexes candidates from weights alone, so missing candidate metadata at inference does not alter its inputs. The initialization ablation concerns how its behavioral training supervision is constructed. We have not retrained LoCAL on targets generated under cold-start or mismatched initialization, which would require regenerating targets for the training corpus and training the encoder again. The results therefore quantify the effect on LoUPE's evaluation, while the retrieval effect of changing the training supervision remains unmeasured.

We will narrow LoUPE's stated scope to behavioral evaluation of adapters with informative starting examples, with community prompts providing the initialization validated here. We will distinguish this setting from LoCAL's retrieval setting, which requires no candidate documentation. We will also add this ablation to the appendix and clarify in Sec. 3.2 and the limitations how initialization quality affects LoUPE's coverage and scores.

AuthorsOpenReview ↗

Response to Reviewer hEMx (3/3)

W2: Scope of LoCAL

Weakness 2: Limited generality of LoCAL. The LoCAL encoder assumes each layer update is a low-rank product of factor matrices (QR/SVD canonicalization, rank standardization, bilinear measurements over paired singular vectors). It therefore cannot handle other parameter-efficient fine-tuning methods that do not produce such weight deltas, such as Textual Inversion. The paper should state this scope explicitly.

We agree that LoCAL's scope should be stated more explicitly. LoCAL is designed for LoRA adapters, and one of our core contributions is learning representations directly from their low-rank weight updates. It does not support methods such as Textual Inversion. We will clarify this scope in the revised manuscript.

Q1: Results on more recent models

Question 1: Could you provide results on more recent models, such as SDXL or DiT-based models (e.g., FLUX)?

Our core contribution is semantic characterization and weight-based LoRA retrieval, whose underlying principles are not specific to SD1.5. We chose SD1.5 for its large public adapter ecosystem and manageable preprocessing costs. Evaluation on SDXL or FLUX would require adapting the encoder, collecting behavioral supervision, and retraining, with substantial additional computation. We have not yet conducted these experiments and acknowledge validation on newer backbones as future work.

Minor point 1: The LoUPE acronym

Minor point 1: The acronym LoUPE is never explicitly expanded; it is only implied by the title of Sec. 3.2. Please define it at first use.

Thank you for pointing this out. LoUPE stands for LoRA Understanding through Prompt Exploration. We will define in Sec. 3.2 of our revision, in the same way as LoCAL.

Reviewer_j6YWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper discusses characterizing, evaluating, and retrieving text-to-image LoRA adapters from a large corpus. This is a growingly important problem with little attention in the literature. On top of the previous work, the paper proposes a new pipeline for constructing adapter-specific prompts, improves upon that lora description features by measuring precision - how much does the lora change the semantic subspace, and finally offers a retrieval model that maps LoRA weights directly into a CLIP-aligned embedding space. The latter is particularly appealing. Finally, the method is evaluated against Stylus, to show inferior results overall, but at a much lower price, making the approach interesting if not already practical.

Critical strengths
  • Interesting and relevant problem setting, with minimal attention in the literature.

  • The Weight-based retrieval is a compelling contribution. It offers a clear potential advantage over previous approaches.

  • A coherent enough evaluation and clear writing.

Critical weaknesses
  • Firstly, it is a big weakness that the paper did not compare against the paper with the closest setting CARLOS. The reason why not is clear, but this lack still makes assessing the quality of the method more difficult.
  • For the experiments that were made, LoCAL improves over Stylus and Stylus+ in several settings, but does not outperform Stylus++. The latter is the only version also using visual information, so it is closer in some sense to the propose one, although again it is clear that they require different resources.
Questions for authors
  • Can CARLoS and LoCAL be compared under a shared evaluation protocol?
  • How should the gap between LoCAL and Stylus++ be interpreted? Could the authors distinguish the effect of the richer information available to Stylus++ from limitations in the learned weight representation? Is there a way to enumerate a trade-off?
  • How well does weight-based retrieval generalize beyond SD1.5 to other diffusion-model architectures and LoRA collections?
4 comments
AuthorsOpenReview ↗

Response to Reviewer j6YW (1/3)

We thank the reviewer for the careful assessment and constructive suggestions on baseline comparisons and the interpretation of our retrieval results.

Our new comparison with CARLoS and metadata-swap experiments with the Stylus family highlight the value of semantic information extracted directly from LoRA-generated images. Recovering these semantics from weights alone is harder, but LoCAL's fast indexing makes this direction worth pursuing.

We detail these experiments and address the reviewer's concerns and questions below.

W1 & Q1: Comparison with CARLoS under a shared protocol

Weakness 1: Firstly, it is a big weakness that the paper did not compare against the paper with the closest setting CARLOS. The reason why not is clear, but this lack still makes assessing the quality of the method more difficult.

Question 1: Can CARLoS and LoCAL be compared under a shared evaluation protocol?

The main reason we did not originally include CARLoS was its substantial indexing cost. Each new LoRA requires generating and CLIP-encoding 280 prompts × 16 seeds = 4,480 images, in addition to a shared base-image bank. The CARLoS paper (§4.1) reports approximately 7 A6000 GPU-hours per LoRA. After training, LoCAL instead takes only 16.3 ms per LoRA on a single L40S GPU for indexing with one forward pass through its weight encoder. CARLoS also focuses primarily on text-to-LoRA retrieval, whereas LoCAL supports both image and text queries.

We now compare both methods under a shared text-to-LoRA retrieval protocol, including the full 1,254-candidate test pool. We retain the original CARLoS paper's 280 indexing prompts and 280 disjoint retrieval templates, but use 2 rather than 16 seeds per prompt, resulting in 560 images per LoRA.

CandidatesMethodR@1 (%)R@5 (%)R@10 (%)MRR
125LoCAL15.35 ± 1.4337.57 ± 1.4651.56 ± 1.480.2693 ± 0.0136
125CARLoS (2 seeds)27.36 ± 1.3347.68 ± 2.1157.17 ± 2.590.3752 ± 0.0154
125Stylus++ (reference)29.81 ± 3.5950.25 ± 4.4859.64 ± 4.830.3863 ± 0.0401
627LoCAL6.09 ± 0.1616.90 ± 0.3124.59 ± 0.310.1250 ± 0.0023
627CARLoS (2 seeds)15.81 ± 0.2930.60 ± 0.5037.84 ± 0.590.2338 ± 0.0035
627Stylus++ (reference)18.71 ± 1.1633.94 ± 0.6541.35 ± 0.780.2526 ± 0.0095
1,254LoCAL4.0711.9017.490.0885
1,254CARLoS (2 seeds)12.6325.0231.570.1907
1,254Stylus++ (reference)14.8728.2235.030.2062

We report per-LoRA averages and include Stylus++ from Table 1. CARLoS outperforms LoCAL but is less accurate than Stylus++. CARLoS's lower performance than Stylus++ may stem from its shared corpus of 280 prompts providing less precise LoRA-specific information than the example images and accompanying prompts used by Stylus++. Both stronger methods use candidate-specific visual evidence, whereas LoCAL tackles the harder task of predicting retrieval semantics from weights. LoCAL's ability to retrieve unseen adapters without candidate-specific generation motivates further work on weight-based retrieval for scalable LoRA search.

AuthorsOpenReview ↗

Response to Reviewer j6YW (2/3)

W2 & Q2: Interpreting the gap between LoCAL and Stylus++

Weakness 2: For the experiments that were made, LoCAL improves over Stylus and Stylus+ in several settings, but does not outperform Stylus++. The latter is the only version also using visual information, so it is closer in some sense to the propose one, although again it is clear that they require different resources.

Question 2: How should the gap between LoCAL and Stylus++ be interpreted? Could the authors distinguish the effect of the richer information available to Stylus++ from limitations in the learned weight representation? Is there a way to enumerate a trade-off?

We tested dependence on candidate information by swapping retrieval inputs between dissimilar LoRAs while keeping weights, candidate IDs, queries, and relevance labels fixed. From the 1,254 test LoRA versions, we selected 137 nonoverlapping pairs (274 LoRAs) whose cosine similarities fall in the bottom 10% for both metadata text (title, description, and trigger words) and example-image representations. We used shuffled greedy matching with seed 42, without using diagnostic queries or retrieval outcomes to select pairs.

Stylus uses titles and descriptions, and Stylus+ adds tags. Stylus++ uses titles, tags, and a GPT summary combining text fields (title, description, tags, and trigger words) with uploader-provided example images and their paired prompts. Partial swaps exchange either source while retaining the other, regenerating summaries with the same GPT instruction. Text-field swaps also replace the retrieval title and tags. Full swaps exchange all retrieval inputs, including existing summaries, without regeneration.

All queries target these 274 LoRAs and are ranked against the same 274 candidates. Stylus variants use 19,152 text queries, while LoCAL uses 19,144 unique-text groups. All methods use the same 57,456 image queries. Metrics average queries within each LoRA, then LoRAs equally. LoCAL represents candidates using weights only, so its results are unchanged by swaps.

Text-to-LoRA retrieval (274 candidates).

MethodConditionR@1 (%)R@5 (%)R@10 (%)MRR@10
StylusOriginal inputs16.8428.8734.350.2197
StylusAll inputs swapped0.090.521.330.0033
Stylus+Original inputs19.4134.3141.950.2587
Stylus+All inputs swapped0.060.420.990.0024
Stylus++Original inputs33.0353.3061.390.4180
Stylus++Text fields swapped31.3249.1056.540.3903
Stylus++Example images + prompts swapped1.353.955.720.0259
Stylus++All inputs swapped0.060.481.040.0026
LoCALWeights only (any condition)11.0727.4339.030.1839

Image-to-LoRA retrieval (274 candidates).

MethodConditionR@1 (%)R@5 (%)R@10 (%)MRR@10
StylusOriginal inputs12.0421.9627.510.1641
StylusAll inputs swapped0.080.721.830.0043
Stylus+Original inputs15.1029.0035.620.2113
Stylus+All inputs swapped0.090.541.360.0032
Stylus++Original inputs28.2947.8856.240.3679
Stylus++Text fields swapped26.4544.0552.230.3410
Stylus++Example images + prompts swapped1.034.136.200.0245
Stylus++All inputs swapped0.070.380.930.0023
LoCALWeights only (any condition)14.4333.5845.210.2274

The partial swaps show that Stylus++'s advantage depends strongly on the semantic information in example images and their paired prompts. Swapping these examples reduces R@1 by about 96% in both modalities and reverses its advantage over LoCAL, whereas swapping text fields reduces R@1 by only 5–7%. Retaining correct examples therefore preserves most of its performance even with incorrect descriptive fields, while correct descriptive fields cannot compensate for strongly mismatched examples.

Together, these comparisons help distinguish the benefit of richer candidate inputs from the capabilities of LoCAL's learned representation. The partial swaps demonstrate Stylus++'s dependence on aligned examples while holding its retrieval method fixed, and the CARLoS results show strong retrieval using generated-image representations. LoCAL is competitive with Stylus and Stylus+ on several metrics, particularly for image retrieval, indicating that it already extracts useful retrieval semantics from weights without explicit descriptive metadata. The gap to Stylus++ and CARLoS suggests that recovering the richer semantics conveyed by visual examples remains a challenge for the current weight representation. Improving this capability is a concrete direction for weight-based retrieval, while retaining the ability to index unseen adapters without candidate metadata or image generation at inference.

AuthorsOpenReview ↗

Response to Reviewer j6YW (3/3)

Q3: Generalization beyond SD1.5

Question 3: How well does weight-based retrieval generalize beyond SD1.5 to other diffusion-model architectures and LoRA collections?

Our core contribution is to characterize and evaluate LoRAs through their semantic effects and learn weight representations for retrieval without candidate metadata or example images at inference. These principles are not inherently specific to SD1.5, which we chose for its large public adapter ecosystem and manageable preprocessing costs. Extending LoCAL to SDXL or FLUX would require adapting the encoder to different layer layouts and weight dimensions, collecting behavioral supervision, and retraining, with substantial additional preprocessing and computation. Our results establish retrieval of held-out SD1.5 adapters within the studied collection. As acknowledged in our limitations, validation on other backbones and collections remains future work.

Dear Reviewer,

The discussion period is nearing its end, and we haven’t yet seen a reply from you. We encourage you to respond to the authors’ comments, especially if you have further questions about points that may affect your assessment.

Thank you for taking the time to engage in the discussion.

Best,

Your AC

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Remarks

We thank all reviewers for their careful reading and constructive feedback. We highlight our main clarifications and additional experiments below.

TL;DR

  • LoRA Scope explores (1) what semantic effects LoRAs exhibit beyond documented examples and (2) how to learn the corresponding representations directly from their weights for efficient retrieval.

  • We propose a pipeline to characterize each LoRA’s semantic effects, a score to assess its training quality, and an encoder to learn semantic representations from its weights.

Main Clarifications

  • In LoRA Scope, we introduced three components for characterizing low-rank adapters and learning their representations. (1) LoUPE targets documented LoRAs, providing a systematic pipeline that iteratively discovers what semantic effects LoRAs exhibit beyond documented examples. (2) LoUPE Score evaluates the LoRA training quality based on the diagnostic prompt set. (3) LoCAL encoder is trained on the LoUPE-enhanced dataset, compressing the whole LoRA weight into a CLIP-space aligned representation. The trained LoCAL encoder enables efficient and metadata-free text/image-to-LoRA retrieval.

Summary of Additional Experiments

  • The results support LoUPE’s adaptive probing as a reproducible way to explore adapter behavior, using community examples as a starting point to uncover a broader range of adapter effects.

  • Expanded comparisons suggest that semantics extracted directly from LoRA-generated images help explain the higher accuracy of Stylus++ and CARLoS. LoCAL’s independence from candidate metadata and efficient single-pass indexing make weight-based retrieval a worthwhile direction for the research community to pursue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.