acceptodds
Under review as a conference paper at ICLR 2027

LoRAHunter: Decoding Semantic Representations from Low-Rank Adapter Weights

Abstract

The rapid growth of community-trained Low-Rank Adaptation (LoRA) adapters creates a need for scalable adapter understanding and retrieval, yet existing systems rely heavily on noisy metadata, example images, or expensive generation-based profiling. We investigate whether LoRA weights can support text-aligned adapter representations learned from metadata and offline generation feedback. We propose LoRAHunter, a structured parameter representation framework that queries low-rank weight updates with learnable probes, projects module-specific responses into a shared token space, and aggregates them into a weight-derived adapter embedding. We align this space with language using noisy metadata as weak supervision and further calibrate it for generation-oriented retrieval using offline rewards. At retrieval time, database-side adapter embeddings are computed from LoRA parameters and support semantic clustering, text-to-LoRA retrieval, and multi-adapter selection. Experiments on Stable Diffusion and Qwen-Image show that the learned parameter representations exhibit meaningful semantic structure, support text-to-LoRA retrieval, and provide a useful signal for generation-oriented adapter selection. These results support supervised weight-derived representations as a complementary signal for adapter analysis and selection.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Diffusion Models, LoRA Retrieval, LoRA Representations

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_s3YqOpenReview ↗
Rating
1
Confidence
4(+1)
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces the LoRAHunter framework, which leverages learnable probes, Transformer aggregation mechanisms, and contrastive learning to derive semantic representations directly from LoRA weights. Furthermore, it incorporates a reward mechanism based on generative tasks to optimize these representations, thereby facilitating "text-to-LoRA" retrieval and multi-adapter selection.

Critical strengths

NONE

Critical weaknesses
  1. The discussion or comparison with existing related work such as W2T, "A LoRA is Worth a Thousand Pictures," and WeightCLIP is insufficient.
  2. In retrieval tasks, the proposed method does not consistently outperform metadata-based baselines. A more thorough comparison with existing weight encoders is required.
  3. The generated results fail to clearly demonstrate whether the learned representations truly capture the actual utility of the LoRA adapters. The paper lacks a direct evaluation based on reward-driven ranking, nor does it assess the compatibility of multi-LoRA combinations.
  4. The set of baselines used in the experiments is too limited, particularly regarding weight-based retrieval. More competitive baselines should be introduced, and fair comparisons must be conducted under identical retrieval and combination settings.
Questions for authors
  1. How does LoRAHunter differ from existing research such as W2T ("A LoRA is Worth a Thousand Pictures") and WeightCLIP?
  2. Why does LoRAHunter not always outperform metadata-based baseline methods in retrieval tasks?
  3. How do the authors demonstrate that the learned representations truly capture the actual functional utility of the LoRA adapters?
  4. Could the authors include more competitive baseline methods specifically those based on weight parameters and evaluate them under the same experimental setup?
5 comments

We thank the reviewer for the constructive feedback and provide our point-by-point responses below.

Re W1 & Q1: Differences from Related Work

Thank you for highlighting these related works. LoRAHunter’s main advance is to connect community LoRA weights with natural-language requests and generation utility, allowing users to describe their desired output and retrieve and select suitable adapters. W2T learns weight representations through task-specific supervision, such as training a classification head to predict 40 predefined CelebA attributes. Its original setup requires additional text alignment to support direct natural-language LoRA search. LoRAHunter aligns weight embeddings with a frozen text encoder, supporting descriptions of objects, scenes, styles, and their combinations. For representation construction, W2T canonicalizes updates using QR–SVD, incurring additional matrix-decomposition overhead. Our learnable probes directly read low-rank updates, with their readout directions learned through semantic supervision. A LoRA is Worth a Thousand Pictures vectorizes LoRA factors and applies PCA with calibration to construct style descriptors, primarily studying style clustering and similarity-based LoRA retrieval. These descriptors do not directly express an adapter’s capabilities in natural language. LoRAHunter combines text alignment with paired generation feedback to learn an index that captures both semantic relevance and generation utility. WeightCLIP aligns full-model weights with dataset image samples, uses target image sets as queries, and evaluates primarily on CNN and ResNet-18 model zoos with a shared architecture. This requires access to corresponding image samples, while the original training data of community LoRAs are often unavailable. We train our encoder using community metadata as weak supervision and offline generation feedback. Once frozen, it can index new LoRAs for the same backbone directly from their weights. This integration of natural-language queries, weight-based indexing, and generation-oriented selection directly serves the needs of expanding community repositories of diffusion adapters. We have added a discussion of these works to the Appendix A.4 (highlighted in blue).

Re W2 & Q2: Why Does LoRAHunter Not Consistently Outperform Metadata-Based Baselines?

As explained in the experimental setup, the information sources used in this retrieval benchmark favor metadata-based methods. The category semantics expressed by the queries and the query–LoRA relevance annotations are derived from adapter metadata through manual annotation or LLM summarization. Metadata CLIP directly accesses titles, descriptions, and tags at retrieval time, including explicit semantic cues aligned with the evaluation target, whereas LoRAHunter must recover these semantics from adapter weights. The advantage of clean metadata on some ranking metrics is therefore consistent with this setup. Meanwhile, LoRAHunter performs better on other measures. Table 1 reports ARI of 0.810 versus 0.799, purity of 0.893 versus 0.876, cosine Silhouette of 0.229 versus 0.205, and an intra–inter similarity gap of 0.347 versus 0.201, comparing LoRAHunter with metadata CLIP. These results support the recovery of meaningful category structure from weights, with stronger performance on these clustering and separation measures. Combined with its ability to index new adapters directly from their weights, LoRAHunter provides a semantic retrieval signal that remains usable when metadata is missing, incomplete, or unreliable.

Re W3 & Q3: How Do We Verify That the Learned Representations Capture the Functional Utility of LoRA Adapters?

For image generation, our primary concern is the quality of the generated outputs. We therefore evaluate LoRA utility directly through generation results. Table 4 in the main paper already shows that the retrieved LoRAs improve text alignment, measured by CLIPScore, and image quality, measured by PickScore and ImageReward. We further compare our retriever with direct retrieval using CLIP metadata embeddings. CLIP-global-Top-1 and Ours-global-Top-1 use CLIP embeddings and our learned weight embeddings, respectively, to retrieve a single LoRA using the full prompt for image generation. Ours-Concept-Top-1 retrieves the top-ranked LoRA for each concept using our weight embeddings and composes the retrieved adapters. Composition(Ours) denotes our full pipeline, while CLIP-Composition replaces our weight-based adapter representations with CLIP embeddings within the same pipeline.

MethodpromptCLIPScore ↑PickScore ↑ImageReward ↑
Base10026.65622.4120.353
CLIP-global-Top-110026.03922.2150.195
Ours-global-Top-110026.70322.4930.402
Ours-Concept-Top-110025.75122.2160.319
CLIP-Composition10026.51822.4410.308
Composition(Ours)10026.92123.0540.554

These empirical results indicate that the learned representations capture both semantic similarity and the functional utility of LoRA adapters. They also show that LoRAs retrieved solely through metadata similarity do not necessarily improve generation quality. LoRA quality varies considerably across the open-source LoRA ecosystem, and textual metadata alone provides limited insight into an adapter’s actual generation behavior. This motivates learning representations that incorporate generation feedback.

Regarding multi-LoRA compatibility, LoRAverse also discusses how composing multiple LoRAs can degrade image quality. We add an experiment that retrieves the top-ranked LoRA for each concept using our encoder and directly composes the retrieved adapters for generation. The results show that naively combining highly ranked LoRAs can introduce compatibility issues and reduce the quality of the generated images. These results have been added to Appendix A.2.6 (highlighted in blue).

Re W4 & Q4: Comparison with Other Weight-Based Methods

Appendix A.2.2 already includes a comparison with a weight-encoding baseline. We additionally implement W2T for comparison. W2T was originally trained with task-specific supervision, such as classification over a predefined set of labels. We adapt it to our setting by retaining its QR–SVD procedure and model architecture while training it with our contrastive learning objective. For a fair comparison, we retrain our model using the same data and training budget. Specifically, both models are trained on 4,000 LoRAs for 20 epochs, providing a controlled comparison of the two encoding approaches. We report retrieval and clustering results below:

MethodHit@1Hit@5Hit@10ARI ↑NMI ↑KNN@5 ↑
W2T0.11670.41670.51670.04580.22180.1505
Ours0.16670.45000.66670.12430.33160.2553

The results show that, under matched training budgets and experimental settings, our method outperforms both the structured weight-encoding baseline and the recent W2T method. Our method also has lower inference overhead. W2T requires matrix decompositions for each LoRA module, adding computational cost. Excluding disk I/O, encoding an SD LoRA takes approximately 0.75 seconds with W2T and 0.5 seconds with our method.These results have been added to Appendix A.2.8 (highlighted in blue).

Dear Reviewer s3Yq,

Thank you again for your comments. We are following up on our detailed response, where we clarified the differences from related work and added controlled experiments to isolate the contributions of weight representations and adapter selection, examine generation utility and composition, and compare additional weight encoders under matched training settings.

Our framework connects LoRA weights with natural-language queries and generation feedback. Its practical value is that users can retrieve relevant LoRAs even when metadata is missing or inaccurate. It also provides a basis for open-source platforms to assess adapter quality. More broadly, interpreting LoRA parameters directly may support future applications in adapter composition and capability customization.

We would appreciate it if you could review these additions and let us know whether they address your main concerns. We would be happy to clarify any remaining questions.

Best regards,
The Authors

Reviewer_YS96OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LoRAHunter, a framework for learning text-aligned representations from LoRA weights. It uses learnable probes and a Transformer to encode module-wise weight updates, aligns the resulting embeddings with textual metadata, and further refines them using offline generation rewards. The learned representations are used for semantic retrieval and multi-adapter selection, with experiments on Stable Diffusion and Qwen-Image.

Critical strengths
  1. Learning semantic representations directly from LoRA weights is an interesting direction, especially when adapter metadata is unreliable or unavailable.
  2. The structured probing mechanism is a reasonable design for handling heterogeneous LoRA modules.
  3. The paper evaluates the approach on two diffusion backbones and includes retrieval, generation, and robustness experiments.
Critical weaknesses
  1. The necessity of the proposed framework is insufficiently justified. The method requires large, separately trained weight encoders (320M/650M parameters), metadata supervision, and substantial offline generation feedback. Given these costs, it is unclear whether learning to predict adapter semantics from weights is preferable to directly profiling adapter behavior and indexing the resulting image features. No compute-matched comparison with such simpler alternatives is provided.
  2. Limited evidence for the effectiveness of the learned weight representations. In Table 2, metadata-based CLIP outperforms LoRAHunter on several retrieval metrics, including MRR and nDCG. The clustering evaluation is limited to 206 adapters from 12 coarse-grained categories. While these results demonstrate feasibility, they provide limited evidence that the proposed architecture learns substantially better semantic representations than simpler alternatives.
  3. The contributions of the weight encoder and downstream selection pipeline are not disentangled. The generation experiments compare complete systems, making it difficult to attribute improvements to the proposed representation rather than preference refinement or heuristic subset selection. In particular, a baseline using metadata CLIP embeddings with the same selection procedure is missing.
  4. The preference refinement evaluation is not sufficiently independent of its training objective. The refinement stage directly optimizes signals derived from CLIPScore and PickScore, which are also used for evaluation. Although ImageReward provides additional evidence, the paper does not compare against simpler reward-aware reranking approaches trained with the same feedback.
  5. The metadata robustness experiment provides limited additional insight. The proposed method does not use metadata at inference, while competing methods rely on deliberately corrupted metadata. Its relative robustness is therefore largely expected by construction and does not establish that the proposed weight encoder is the most effective solution to missing or noisy metadata.
Questions for authors
  1. Given the substantial offline generation and training costs, how does LoRAHunter compare with directly profiling LoRAs using a small set of fixed prompts and indexing the resulting image embeddings, under matched computational budgets?
  2. How much of the generation improvement is attributable specifically to the learned weight representations? What happens when metadata CLIP embeddings are used with exactly the same retrieval and subset-selection pipeline?
  3. Can simpler structured weight encoders or reward-aware rerankers achieve comparable performance under the same supervision and computation budgets?
  4. What is the total computational cost of building and updating the adapter index, including preference-data generation? How does this scale to substantially larger LoRA repositories?
5 comments

We thank the reviewer for the constructive feedback and provide our point-by-point responses below.

Re W1: Necessity for the Proposed Framework

LoRAHunter aims to learn reusable semantic representations of LoRA weights, enabling natural-language retrieval and adapter selection in growing community repositories with varying metadata quality. Indexing generated images characterizes an adapter through its outputs under particular prompts and random seeds, so the behavior captured depends on these probing conditions. A small set of fixed prompts may provide limited coverage of an adapter’s effects across different objects, scenes, and style combinations. LoRA weights encode the updates to model computation, motivating a shared mapping from these updates to language semantics. Once the encoder is frozen, new LoRAs can be mapped directly from their weights into a shared semantic space for text-based retrieval and composition selection. The practical value of our framework is that users can retrieve relevant LoRAs even when metadata is missing or inaccurate. It also provides a basis for open-source platforms to assess adapter quality. More broadly, interpreting LoRA parameters directly may support future applications in adapter composition and capability customization.

Re Q1: Comparison with Image-Based Indexing

We consider directly indexing embeddings of LoRA-generated images problematic, as these embeddings may primarily reflect the input prompt rather than the LoRA’s contribution. To isolate the adapter’s effect, we conduct an embedding-difference analysis. We subtract the embedding of an image generated by the base model from that of the corresponding LoRA-generated image, using the same prompt and random seed. Intuitively, this difference can capture aspects of the LoRA’s functional effect. For each LoRA, we generate paired images using five fixed prompts and average their embedding differences. We report retrieval performance on a test set of 800 LoRAs below.

MethodHit@1Hit@5Hit@10nDCG@10
image prompt0.100.300.4400.143
ours0.4200.6600.8200.465

Here are the results of the cost comparison:

MethodIndex build (GPU-h)Index storageRaw image storageOne-time training
image prompt6.371≈4.0 MB≈1.9 GBNone
ours0.32 (Including loading LoRA from disk)≈8.1 MBNone96 GPU-h (full training)

These results show that our encoder provides more effective retrieval than image-based indexing. Using more profiling prompts may improve image-based retrieval, but it also increases the cost of generating images. Although our method incurs an upfront training cost, this is a one-time investment over the full adapter corpus. Once trained, the encoder can index new LoRAs at very low cost. Image-based indexing, however, requires additional image generation for every new LoRA, making it impractical for community repositories containing tens of thousands of LoRAs. Retaining the generated images also incurs substantial storage costs. These results have been added to Appendix A.2.7 (highlighted in blue).

Re Q2 & W3: How Much of the Generation Quality Improvement Is Attributable to the Learned Weight Representations?

To disentangle the contributions of the LoRA selection pipeline and weight representations to generation performance, we conducted additional ablation studies. The results are summarized below:

MethodpromptCLIPScore ↑PickScore ↑ImageReward ↑
Base10026.65622.4120.353
CLIP-global-Top-110026.03922.2150.195
Ours-global-Top-110026.70322.4930.402
Ours-Concept-Top-110025.75122.2160.319
CLIP-Composition10026.51822.4410.308
Composition(Ours)10026.92123.0540.554

CLIP-global-Top-1 and Ours-global-Top-1 use CLIP embeddings and our learned weight embeddings, respectively, to retrieve a single LoRA using the full prompt for image generation. Ours-Concept-Top-1 retrieves the top-1 LoRA for each concept using our weight embeddings and composes the retrieved LoRAs. Composition(Ours) denotes our full pipeline, while CLIP-Composition replaces our weight-based adapter representations with CLIP embeddings within the same pipeline. The results show that our encoder captures information about the functional utility of LoRA adapters. They also indicate that naively composing LoRAs can introduce compatibility issues and degrade image quality, while our subset-selection procedure mitigates this risk. These results have been added to Appendix A.2.6 (highlighted in blue).

Re W2 & Q3: Comparison with Other Weight Encoders

Appendix A.2.2 already includes a comparison with a weight-encoding baseline. We additionally implement W2T[1] for comparison, We have added a discussion of these works to the Appendix A.4 (highlighted in blue). W2T was originally trained with task-specific supervision, such as classification over a predefined set of labels. We adapt it to our setting by retaining its QR–SVD procedure and model architecture while training it with our contrastive learning objective. For a fair comparison, we retrain our model using the same data and training budget. Specifically, both models are trained on 4,000 LoRAs for 20 epochs, providing a controlled comparison of the two encoding approaches. We report retrieval and clustering results below:

MethodHit@1Hit@5Hit@10ARI ↑NMI ↑KNN@5 ↑
W2T0.11670.41670.51670.04580.22180.1505
Ours0.16670.45000.66670.12430.33160.2553

The results show that, under matched training budgets and experimental settings, our method outperforms both the structured weight-encoding baseline and the recent W2T method. These results have been added to Appendix A.2.8 (highlighted in blue).

[1]W2T: LoRA Weights Already Know What They Can Do

Re W4 & Q3: Comparison with a Lightweight Reward-Aware Reranker

We train a simple reward-aware reranker using the same feedback. It takes the weight embeddings from the Stage 1 LoRA encoder and text-query embeddings as inputs and predicts a prompt–adapter score, using the pairwise logistic loss �0��0� . The reranker does not outperform direct retrieval using our encoder.

MethodpromptCLIPScore ↑PickScore ↑ImageReward ↑
Base10026.65622.4120.353
reranker-top-110025.63122.1350.145
Ours-global-Top-110026.70322.4930.402

The reranker learns scores from fixed Stage 1 representations and is therefore limited to the utility information retained in those embeddings. Our preference-weighted contrastive learning directly updates the LoRA encoder, giving high-reward pairs greater influence on representation learning and incorporating generation preferences into adapter–text similarity. The resulting cached index vectors support preference-aware retrieval and remain reusable for subsequent subset selection and broader downstream tasks, providing utility beyond a standalone scoring model. These results have been added to Appendix A.2.9 (highlighted in blue).

Re Q4: Computational Cost Analysis

As reported in Appendix A.2.4, the encoder takes approximately 24 hours to train on four NVIDIA A100 GPUs. Preference data generation and scoring are performed during training preparation and constitute a one-time investment. In our SD experiments, this stage uses 20K prompts to generate approximately 150K images and takes about one day on four NVIDIA A100 GPUs. Once the encoder is frozen, building an index of 5,000 LoRAs takes approximately two hours on a single GPU, including disk loading. Adding a new LoRA for the same backbone requires only loading its weights, encoding it once, and inserting its embedding into the index. Encoding takes approximately 0.5 seconds per LoRA, excluding disk loading, resulting in a small incremental overhead. Adding new LoRAs requires neither regenerating preference data nor retraining the encoder. As the repository grows, loading and encoding work scales approximately linearly with the number of adapters to be indexed. Updates process only new or modified adapters, while the one-time training and preference-data preparation costs are amortized across an expanding repository.

Re W5: Value of the Metadata Robustness Evaluation

Community LoRAs are contributed by different developers, resulting in varying completeness and information quality in their titles, descriptions, and tags. Missing descriptions and uninformative titles are practical deployment challenges. Our controlled perturbations quantify how these field-level issues affect downstream generation. The weight index’s invariance to metadata perturbations follows from its design, while the experiment measures the practical impact of relying on textual metadata. The comparison reported in our response to Q2 & W3 also shows that CLIP metadata embeddings do not outperform weight-based encoding. Together, these results demonstrate that weight indexing provides a competitive semantic retrieval signal while removing inference-time dependence on the quality of each LoRA’s text fields. The perturbation experiment adds deployment-relevant evidence under realistic repository conditions, further supporting the practical advantages of our approach.

Dear Reviewer YS96,

Thank you again for your constructive feedback. We have posted a point-by-point response with additional controlled ablations, comparisons with image-based indexing and reward-aware reranking, and a detailed computational cost analysis.

We would appreciate it if you could review these additions and let us know whether they address your main concerns. We would be happy to clarify any remaining questions.

Best regards, The Authors

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

This paper proposes LoRAHunter, which encodes LoRA weights using learnable probes and a Transformer, and learns text-aligned weight representations through metadata-supervised contrastive learning and generation-based preference refinement, reducing reliance on external metadata during retrieval. Building on these representations, the method further enables multi-LoRA selection through concept decomposition and subset scoring. Experiments demonstrate that the learned weight representations capture meaningful semantic information and show advantages in generation quality and robustness to metadata degradation.

Critical strengths

1. Clear research motivation and practical value of weight-based semantic representations. By aligning LoRA weights with textual semantics, the proposed method enables LoRA retrieval without relying on external metadata at inference time. It learns semantically meaningful representations directly from LoRA weights and achieves competitive performance across multiple evaluation metrics, demonstrating practical value, especially when metadata is missing or unreliable.

2. Well-structured experimental evaluation from multiple perspectives. Semantic clustering and retrieval experiments show that the learned LoRA weight representations can be effectively aligned with textual semantics and support text-to-LoRA retrieval. Image generation and metadata degradation experiments further demonstrate their usefulness for LoRA selection and robustness to unreliable metadata, providing support for the method's practical value.

Critical weaknesses

1. Insufficient evaluation of the independent contribution of weight representations to generation performance. The final generation performance is influenced by weight representations, preference refinement, and multi-LoRA selection. Although Table 4 validates the effect of preference refinement, the paper lacks a direct comparison between LoRAHunter weight representations and CLIP metadata representations under the same retrieval and composition strategy. Using only LoRA weights for retrieval is an important feature of this work, but this comparison would help show whether weight representations also improve generation quality.

2. Lack of direct evaluation of the multi-LoRA selection strategy. The proposed method selects multiple LoRAs through concept decomposition and subset scoring. Although Table 13 in the appendix analyzes generation performance across different numbers of concepts, the paper lacks a direct comparison with global LoRA selection based only on the full prompt. Such a comparison is important for determining whether the improvements in generation performance actually come from concept decomposition and multi-LoRA composition.

Questions for authors

Could the authors provide more details on how the metadata degradation benchmark is constructed, along with examples before and after perturbation? The paper simulates metadata corruption by randomly removing description fields or replacing titles, but it is unclear how well these perturbations reflect real-world metadata issues. Different perturbation strategies may affect the performance of metadata-based baselines, further clarification of these settings would help strengthen the reliability of the robustness evaluation.

1 comment
AuthorsOpenReview ↗

Contributions of Weight Representations and Details of metadata degradation benchmark

We thank the reviewer for the constructive feedback and provide our point-by-point responses below.

Re W1 & W2: Contributions of Weight Representations and Multi-LoRA Selection

To disentangle the contributions of the LoRA selection pipeline and weight representations to generation performance, we conducted additional ablation studies. The results are summarized below:

MethodpromptCLIPScore ↑PickScore ↑ImageReward ↑
Base10026.65622.4120.353
CLIP-global-Top-110026.03922.2150.195
Ours-global-Top-110026.70322.4930.402
Ours-Concept-Top-110025.75122.2160.319
CLIP-Composition10026.51822.4410.308
Composition(Ours)10026.92123.0540.554

CLIP-global-Top-1 and Ours-global-Top-1 use CLIP embeddings and our learned weight embeddings, respectively, to retrieve a single LoRA using the full prompt for image generation. Ours-Concept-Top-1 retrieves the top-1 LoRA for each concept using our weight embeddings and composes the retrieved LoRAs. Composition(Ours) denotes our full pipeline, while CLIP-Composition replaces our weight-based adapter representations with CLIP embeddings within the same pipeline. The results show that our encoder captures information about the functional utility of LoRA adapters. They also indicate that naively composing LoRAs can introduce compatibility issues and degrade image quality, while our subset-selection procedure mitigates this risk. These results have been added to Appendix A.2.6 (highlighted in blue).

Re Q1: Details on how the metadata degradation benchmark is constructed

We use the same LoRA candidate pool. For each LoRA record, the description has a 50% probability of being removed, and the title has a 50% probability of being replaced with an unrelated term or a semantically uninformative identifier. These probabilities apply at the field level, allowing some records to retain their complete original metadata. We evaluate generation performance on 250 COCO prompts. Stylus also identifies poor-quality metadata as a practical issue for LoRAs on open-source platforms. It reports a source collection of over 100K LoRAs on CivitAI and a curated StylusDocs dataset of approximately 75K LoRAs, following filtering that includes the exclusion of NSFW entries. We provide examples before and after perturbation below. These results have been added to Appendix A.3 (highlighted in blue).

LoRAFieldBeforeAfter
LoRA_1titleWaist Slider (Microwaist)Waist Slider (Microwaist)
LoRA_1description“This LoRA focuses on modifying the waist size of characters, creating an extremely narrow waist…”" "
LoRA_2titleAmber (Genshin Impact)untitled_model_8082
LoRA_2description"This LoRA is designed to generate images of the character Amber from the game Genshin Impact. The adapter accurately captures Amber's distinctive features, including her brown eyes, ....""This LoRA is designed to generate images of the character Amber from the game Genshin Impact. The adapter accurately captures Amber's distinctive features, including her brown eyes, ...."

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.