single-lora style adaptation

This work addresses a central challenge in parameter-efficient fine-tuning: identifying the optimal placement of adapters to achieve peak performance with minimal parameters. The authors propose PAGE, a metric based on initial gradient energy analysis, which reveals that adaptation effects are highly concentrated in the down-projection modules of shallow feed-forward networks. Leveraging this insight, they introduce DomLoRA—a method that deploys a single LoRA adapter exclusively in this dominant module. This study is the first to demonstrate the existence, architectural dependency, and task stability of such a dominant adaptation module, establishing a new paradigm for efficient fine-tuning. Experiments show that DomLoRA, using only ~0.7% of the parameters of standard LoRA, consistently outperforms it across diverse tasks—including instruction following, mathematical reasoning, code generation, and multi-turn dialogue—and further enhances the effectiveness of other LoRA variants.

Existing image-guided style transfer methods suffer from low structural fidelity and difficulty in balancing content and style during inference, primarily due to the tight coupling between content and style representations and conflicts arising from multiple adapter modules. To address these limitations, this work proposes AnyStyle, a novel framework that abandons multi-adapter designs and instead introduces a single LoRA module to unify style modeling. Leveraging the internal self-attention mechanisms of pre-trained diffusion models, AnyStyle extracts content structure guidance signals without requiring additional training. This approach significantly enhances controllability, stability, and computational efficiency during style transfer. While maintaining competitive quantitative performance, AnyStyle achieves more accurate structure preservation and higher-quality artistic stylization compared to existing methods.

This work addresses the challenge in personalized image generation where existing LoRA composition methods struggle to balance content fidelity and style consistency, often resulting in entangled content-style representations, weak controllability, and unstable fusion. To overcome these limitations, the authors propose a training-free decoupled fusion framework that separates content and style subspaces through rank-constrained fine-tuning. They introduce a prompt-guided multi-branch expert encoder to enable semantically controllable adapter aggregation and incorporate a classifier-free temporal coherence guidance mechanism to enhance generation stability. This approach achieves, for the first time, retraining-free disentangled LoRA fusion that simultaneously preserves high-fidelity content and supports flexible semantic control, significantly outperforming current state-of-the-art methods.

To address performance degradation and concept forgetting caused by rapid switching among multiple LoRA modules, this paper proposes Sparse High-Rank Adapters (SHiRA)—a parameter-efficient, zero-inference-overhead adaptation paradigm. SHiRA fine-tunes only 1–2% of the base model’s parameters, enabling millisecond-scale adapter switching via structured sparse weight updates and fusion-aware training, while supporting collaborative multi-adapter fusion. Theoretical analysis elucidates how high sparsity enhances multi-task synergy. SHiRA is fully compatible with mainstream LLMs and LVMs without requiring modifications to inference engines. Experiments demonstrate that SHiRA consistently outperforms LoRA across multiple large models: it reduces GPU memory peak usage by 16%, accelerates CPU loading speed by 5–16×, and achieves training efficiency comparable to LoRA.

Real-time fusion of content and style LoRAs in personalized image generation remains challenging due to high computational overhead of existing optimization-based methods, rendering them unsuitable for edge devices. Method: This paper proposes a hypernetwork-based dynamic LoRA fusion paradigm that bypasses iterative optimization; instead, a lightweight hypernetwork directly predicts optimal merging weights for millisecond-level, high-fidelity synthesis. Contribution/Results: We introduce an MLLM-driven joint content-style evaluation protocol to overcome biases inherent in conventional metrics, and adopt a cross-content–style generalization training strategy to significantly enhance model generalizability. Experiments demonstrate that our method accelerates fusion by over 4,000× compared to state-of-the-art optimization approaches, while achieving new benchmarks in both content and style fidelity—rigorously validated via automated MLLM assessment and human evaluation.

This work proposes D2-LoRA, a parameter-efficient fine-tuning method designed for scenarios with limited training data and computational resources, where existing approaches often struggle to balance performance, stability, and mergeability. D2-LoRA integrates signed low-rank residual updates that encode both directional and differential information, complemented by column-wise norm projection to constrain weight updates. This yields an adapter architecture that is highly accurate, exhibits low training variance, and supports algebraic merging without introducing inference latency. Evaluated across eight question answering and reading comprehension benchmarks, D2-LoRA achieves an average accuracy of 76.4%, outperforming LoRA by 2.2 percentage points, while reducing training instability by 36%. The merged model demonstrates a 1.91× increase in inference throughput with negligible numerical degradation of approximately 0.03 percentage points.

This work addresses the parameter inefficiency of conventional LoRA-based continual learning, where each task is assigned a separate adapter despite significant low-rank redundancy across tasks. The authors propose LiteLoRA, a novel approach that leverages a plug-in gating mechanism and subspace overlap analysis to selectively reuse existing adapters or instantiate new ones in a dynamic, task-aware manner. By identifying and exploiting shared low-rank subspaces among tasks, LiteLoRA achieves competitive or superior performance compared to state-of-the-art methods on standard continual learning benchmarks, while reducing the number of active adapters by 20% to 70%, thereby substantially improving parameter efficiency.

Existing LoRA fusion methods for image generation often suffer from content shift and detail degradation, making it challenging to achieve high-quality multi-style transfer efficiently. This work introduces frequency-domain analysis into LoRA fusion for the first time, proposing a training-free dynamic switching mechanism that adaptively selects adapters based on frequency-domain importance while preserving content consistency through a semantic alignment strategy. The proposed approach substantially reduces the training cost of customized generation and effectively mitigates content drift and detail loss in complex tasks involving multiple objects and styles, all while maintaining high fidelity and generation quality.

Existing image stylization methods struggle to balance inference efficiency and style fidelity: adapter-based approaches often lose style specificity, while personalization techniques such as LoRA require per-style training. This work proposes i2L, a framework that, for the first time, fully feed-forward generates stylized LoRA weights—leveraging an image encoder, learnable LoRA queries, and a compact decoder head to predict text-to-image model LoRA parameters directly from one or multiple reference images, enabling optimization-free instant style instantiation. The method supports asymmetric classifier-free guidance, multi-style fusion, and seamless integration with controllable generation modules, effectively suppressing content copying while preserving prompt alignment. Experiments demonstrate that i2L significantly outperforms existing approaches on Z-Image, FLUX.2, and Hidream-O1, achieving consistent improvements in style fidelity, prompt adherence, and perceptual quality.