arXiv:2511.11051v1 Announce Type: new 
Abstract: Low-Rank Adaptation (LoRA) fusion has emerged as a key technique for reusing and composing learned subject and style representations for controllable generation without costly retraining. However, existing methods rely on weight-based merging, where one LoRA often dominates the other, leading to interference and degraded fidelity. This interference is structural: separately trained LoRAs occupy low-rank high-dimensional subspaces, leading to non-orthogonal and overlapping representations. In this work, we analyze the internal structure of LoRAs and find their generative behavior is dominated by a few principal directions in the low-rank subspace, which should remain free from interference during fusion. To achieve this, we propose Null Space Projection LoRA (NP-LoRA), a projection-based framework for LoRA fusion that enforces subspace separation to prevent structural interference among principal directions. Specifically, we first extract principal style directions via singular value decomposition (SVD) and then project the subject LoRA into its orthogonal null space. Furthermore, we introduce a soft projection mechanism that enables smooth control over the trade-off between subject fidelity and style consistency. Experiments show NP-LoRA consistently improves fusion quality over strong baselines (e.g., DINO and CLIP-based metrics, with human and LLM preference scores), and applies broadly across backbones and LoRA pairs without retraining.

يقدم المقال NP-LoRA، وهو إطار جديد لدمج التكيف منخفض الرتبة (LoRA) يعالج مشكلة التداخل في الأساليب الحالية. غالبًا ما تؤدي الدمج القائم على الوزن إلى هيمنة LoRA واحد على الآخر، مما يؤدي إلى تدهور الدقة. يستخدم NP-LoRA نهجًا قائمًا على الإسقاط للحفاظ على فصل الفضاءات الفرعية، مما يحسن جودة الدمج من خلال منع التداخل الهيكلي بين الاتجاهات الرئيسية.

El artículo presenta NP-LoRA, un nuevo marco para la fusión de Adaptación de Bajo Rango (LoRA) que aborda el problema de la interferencia en los métodos existentes. La fusión basada en pesos a menudo conduce a que un LoRA domine al otro, lo que resulta en una degradación de la fidelidad. NP-LoRA utiliza un enfoque basado en proyección para mantener la separación de subespacios, mejorando así la calidad de la fusión al prevenir la interferencia estructural entre las direcciones principales.

L'article présente NP-LoRA, un nouveau cadre pour la fusion d'adaptation à faible rang (LoRA) qui aborde le problème d'interférence dans les méthodes existantes. La fusion basée sur le poids conduit souvent à la domination d'un LoRA sur un autre, entraînant une dégradation de la fidélité. NP-LoRA utilise une approche basée sur la projection pour maintenir la séparation des sous-espaces, améliorant ainsi la qualité de la fusion en empêchant l'interférence structurelle entre les directions principales.

The article introduces NP-LoRA, a novel framework for Low-Rank Adaptation (LoRA) fusion that addresses the issue of interference in existing methods. Traditional weight-based merging often leads to one LoRA dominating another, resulting in degraded fidelity. NP-LoRA utilizes a projection-based approach to maintain subspace separation, thereby enhancing the quality of fusion by preventing structural interference among principal directions.

NP-LoRA: Null Space Projection Unifies Subject and Style in LoRA Fusion

Was this article worth reading? Share it

Ready to build your own newsroom?